Railway track real scene three-dimensional modeling method and device based on spatial constraint filtering

Abstract

The present application relates to railway track digital modeling and three-dimensional visualization technical field, provide a kind of railway track real scene three-dimensional modeling method and device based on space constraint filtering.The laser scanner, camera and positioning and pose system are integrated at the bottom of operating train to carry out dynamic acquisition along the line, based on standard track profile three-dimensional pipeline area, the point cloud is filtered out noise based on space constraint, and according to pose information, effective point cloud is projected to image plane multi-frame weighted fusion to generate true color point cloud, and three-dimensional model is output by triangulation, solve the problems of long modeling cycle, long time-consuming of point cloud noise reduction, easy to delete features, low fidelity of traditional segmented scanning, with the advantages of realizing rapid construction of full-line model, extremely fast noise reduction and high fidelity fusion, improve modeling efficiency and fidelity.

Description

Technical Field

[0001] This application relates to the field of railway track digital modeling and 3D visualization technology, specifically to a method and apparatus for real-world 3D modeling of railway tracks based on spatial constraint filtering. Background Technology

[0002] As the lifeblood of the national economy, railways have seen continuous growth in their operational mileage, creating an urgent need for digital operation and maintenance of infrastructure. Traditional track inspection relies mainly on manual track patrols and track inspection vehicles, which suffers from low efficiency, long cycles, and significant human error. With the advancement of the "Digital Railway" strategy, constructing rapid 3D realistic models of railway tracks has become a fundamental step in realizing digital twins. However, railway lines are long and strip-shaped, with complex terrain along the routes, requiring 3D modeling to achieve an optimal balance between acquisition speed, millimeter-level accuracy, and full automation. Each existing technological approach has its advantages and disadvantages: for example, traditional static photogrammetry has limited accuracy, making it difficult to meet the needs of rapid and high-precision modeling of railway lines; although ground laser scanning has high accuracy, its operation requires line closure work, resulting in extremely low efficiency in producing a 3D model of the entire line; mobile laser scanning can balance efficiency and accuracy, and is gradually becoming the mainstream carrier for rapid modeling in industrial engineering, but its application on railway tracks is still relatively limited. Existing laser scanning mostly adopts static station setup, requiring frequent station setup and relocation, and usually requires track maintenance windows. It takes weeks or even months to complete the modeling of hundreds of kilometers of track, which cannot meet the needs of rapid modeling of the entire line; in addition, when stitching together the track models generated by segmented scanning, it is difficult to unify the coordinates of the virtual models corresponding to different times and spaces, making it difficult to form a complete 3D model of the entire line; at the same time, there are a large number of noise points in the railway environment (such as adjacent trains, trees, fences, birds, etc.), and traditional statistical filtering or radius filtering only relies on local point cloud density, which cannot utilize the standard prior knowledge of track geometry, resulting in long noise reduction time (several seconds for millions of points) and easy to mistakenly delete rail edge features or miss near-rail noise.

[0003] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0004] The purpose of this application is to provide a method and device for 3D modeling of railway tracks based on spatial constraint filtering. It has the advantages of efficient and rapid construction of a real-scene model of the entire railway track, rapid noise reduction of point clouds and high-fidelity fusion of true color, thereby improving modeling efficiency and model realism.

[0005] Firstly, this embodiment relates to a method for 3D modeling of railway tracks based on spatial constraint filtering, including: Calibrate the extrinsic parameters of the 3D laser scanner, camera, and positioning and attitude determination system to the carrier coordinate system; During train operation, laser point cloud, images, and positioning and attitude determination data along the track are collected synchronously based on the second pulse signal of the global navigation satellite system. The positioning and attitude data are combined for navigation calculation to obtain the train's position and attitude information at continuous moments, and the laser point cloud is converted to the world coordinate system based on the extrinsic parameters; Based on a pre-constructed standard track contour 3D pipeline region, spatial constraint filtering is applied to the converted laser point cloud to remove noise points located outside the 3D pipeline region, thus obtaining an effective point cloud; Based on the pose information and extrinsic parameters, the effective point cloud is projected onto the image plane with the corresponding timestamp, pixel color values ​​are extracted and multi-frame weighted fusion is performed to generate a true color point cloud; Based on the true-color point cloud, triangulation is performed, and after mesh optimization, hole repair and texture mapping, a real-world 3D model of the railway track is output.

[0006] Furthermore, the extrinsic parameters include: The rotation matrix and translation vector of the 3D laser scanner to the carrier coordinate system; The intrinsic parameter matrix of the camera, and the rotation matrix and translation vector of the camera to the carrier coordinate system.

[0007] Furthermore, the synchronous acquisition of laser point cloud, imagery, and positioning and attitude data along the track includes: Using the second pulse signal of the global navigation satellite system as a reference, a synchronous trigger signal is generated by frequency division through a field-programmable gate array to control the three-dimensional laser scanner and the camera to collect data synchronously at a preset frequency, so that the data of each sensor has a unified timestamp and the time error is less than the preset duration.

[0008] Furthermore, the step of performing integrated navigation calculation on the positioning and attitude determination data includes: By employing tightly coupled Kalman filtering and fusing data from the Global Navigation Satellite System, the Inertial Measurement Unit, and the Odometer Pulse Data, the position, speed, and attitude angle of the train at continuous time points are calculated.

[0009] Furthermore, the spatial constraint filtering of the converted laser point cloud includes: During periods when no trains are running, a static scanner is used to acquire standard track point cloud samples. Rail cross-sections are extracted along the longitudinal direction of the track at preset intervals. A random sampling consensus algorithm is used to fit the standard rail profile and expand it by a preset distance to generate a two-dimensional profile. Based on the route design parameters, the two-dimensional contour is spatially transformed and interpolated along the longitudinal direction of the track with a preset step size to generate a continuous three-dimensional pipeline region. The current measured point cloud is iteratively registered with the predefined contour of the corresponding mileage, and the translation and rotation are calculated. If the median registration error is greater than the preset threshold, the contour parameters are updated. Based on the mileage coordinates of each point, the corresponding contour section is retrieved. It is determined whether the registered and compensated points are located within the three-dimensional pipeline area. Points located outside the area are removed, while points located within the area and those less than a preset distance from the boundary are retained. The retained valid point cloud is then downsampled using voxel downsampling with a preset voxel side length.

[0010] Furthermore, the projection of the effective point cloud onto the image plane corresponding to the timestamp includes: Based on the extrinsic parameters and pose information, the effective point cloud is transformed from the world coordinate system to the camera coordinate system; Using the camera's intrinsic parameter matrix, three-dimensional points in the camera coordinate system are projected onto the pixel plane to obtain the corresponding pixel coordinates; If the pixel coordinates are within the image range, then the RGB color value of that pixel is extracted.

[0011] Furthermore, before projecting the 3D points in the camera coordinate system onto the pixel plane, an octree spatial index is constructed, and the candidate image frame corresponding to each valid point is quickly located based on the octree.

[0012] Furthermore, the multi-frame weighted fusion includes: for the same physical point covered by multiple frames of images, calculating the distance weight based on the depth value of the point in the camera coordinate system, and performing a weighted average of the RGB color values ​​extracted from each frame according to the distance weight.

[0013] Furthermore, the triangulation based on true-color point clouds includes: The initial mesh is generated using a parallel Delaunay triangulation algorithm, and a threshold for the maximum triangle side length is set. Detect boundary edges and fill holes using the minimum angle method, and simplify the mesh using a quadratic error metric algorithm; The pixel coordinates of the mesh vertices are normalized to texture coordinates based on the width and height of the texture image, generating a textured real-world 3D model and outputting it in a preset format. At the same time, a metadata file containing the coordinate system, accuracy, and mileage range is generated.

[0014] Secondly, this application also proposes a 3D modeling device for railway tracks based on spatial constraint filtering, used to perform the method described in the first aspect, including: The calibration module is used to calibrate the extrinsic parameters of the 3D laser scanner, camera, and positioning and attitude determination system to the carrier coordinate system; The acquisition module is used to synchronously acquire laser point cloud, imagery, and positioning and attitude data along the track during train operation, based on the second pulse signal of the Global Navigation Satellite System. The navigation module is used to perform combined navigation calculations on the positioning and attitude data, obtain the position and attitude information of the train at continuous time intervals, and convert the laser point cloud to the world coordinate system based on the extrinsic parameters. The filtering module is used to perform spatial constraint filtering on the converted laser point cloud based on a pre-constructed standard track contour three-dimensional pipeline region, to remove noise points located outside the three-dimensional pipeline region, and obtain an effective point cloud; The fusion module is used to project the effective point cloud onto the image plane with the corresponding timestamp based on the pose information and extrinsic parameters, extract pixel color values ​​and perform multi-frame weighted fusion to generate a true color point cloud; The reconstruction module is used to perform triangulation based on the true-color point cloud, and after mesh optimization, hole repair and texture mapping, output a real-world 3D model of the railway track.

[0015] As can be seen from the above, the railway track real-scene 3D modeling method and device provided in this application, based on spatial constraint filtering, constructs a mobile sensing source by integrating a 3D laser scanner, camera and positioning and attitude determination system on an operating train to collect dynamic data along the entire line. Based on the pre-constructed standard track contour 3D pipeline region, spatial constraint filtering is applied to the converted laser point cloud to remove noise. Based on the pose information and extrinsic parameters, the effective point cloud is projected onto the image plane and multi-frame weighted fusion is performed to generate a true-color point cloud. This solves the problems of low efficiency in the rapid generation of complete models, long time consumption for point cloud noise reduction and low model fidelity in traditional railway track modeling. It has the advantages of being able to realize efficient and rapid construction of real-scene models of the entire railway track, rapid noise reduction of point clouds and high-fidelity fusion of true color, thus improving modeling efficiency and model realism. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a flowchart illustrating the steps of the method for real-world 3D modeling of railway tracks based on spatial constraint filtering disclosed in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the steps of spatial constraint filtering on the converted laser point cloud as disclosed in an embodiment of the present invention. Figure 3 This is a schematic diagram of the functional framework of the rapid 3D modeling system for railway tracks based on spatial constraint filtering disclosed in an embodiment of the present invention. Figure 4This is a schematic diagram of the operation process of the railway track real-scene 3D modeling system based on spatial constraint filtering disclosed in the embodiments of the present invention; Figure 5 This is a schematic diagram of the structure of the railway track real-scene 3D modeling device based on spatial constraint filtering disclosed in the embodiments of the present invention. Detailed Implementation

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments belong; the terminology used herein and in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit these embodiments; the terms "comprising" and "having," and any variations thereof, in the specification of these embodiments and the foregoing drawings, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification of these embodiments and the foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0019] The implementation details of the technical solution in this embodiment are described in detail below: Firstly, this application proposes a method for 3D modeling of railway tracks based on spatial constraint filtering, such as... Figure 1 As shown, the method includes: S101, calibrating the extrinsic parameters of the 3D laser scanner, camera, and positioning and attitude determination system to the carrier coordinate system; The extrinsic parameters include: the rotation matrix and translation vector of the 3D laser scanner to the carrier coordinate system; the intrinsic parameter matrix of the camera; and the rotation matrix and translation vector of the camera to the carrier coordinate system.

[0020] Specifically, a 3D laser scanner, a high-definition action camera, and a positioning and attitude determination system (POS) are rigidly integrated into the work platform (the bottom of the operating train), and the coordinate systems of the 3D laser scanner, camera, and platform are calibrated. The coordinate system calibration method mainly includes the coordinate calibration of the 3D laser scanner and the high-definition action camera. Using the platform coordinates as a reference, the coordinate systems of the 3D laser scanner and the high-definition action camera are calibrated together to ensure that they have a unified reference coordinate system (enabling spatial alignment). The specific calibration method is as follows: Transformation relationship from 3D laser scanner to carrier coordinate system:

[0021] Among them, P L P represents the coordinates of a point in the coordinate system of a 3D laser scanner. B R represents the coordinates of a point in the carrier coordinate system. L B and t L BThe rotation matrix and translation vector from the 3D laser scanner to the carrier coordinate system.

[0022] The relationship between camera pixels after intrinsic parameter correction and their transformation to the carrier coordinate system:

[0023] Where K is the camera intrinsic parameter matrix, P u Let R be the homogeneous coordinates of the image pixels (u,v,1)^T. C B and t C B These represent the rotation matrix and translation vector from the camera to the carrier coordinate system, respectively. During calibration, the known coordinates P of the target in the carrier coordinate system are used. B and its corresponding homogeneous pixel coordinates P in the image. u As input, the rotation matrix R is iteratively optimized by minimizing the reprojection error or using the direct linear transformation method. C B Translation vector t C B .

[0024] The calculated R L B t L B R C B t C B The parameters are written into the system's configuration file to complete the extrinsic parameter calibration of the 3D laser scanner, camera, and positioning and orientation system to the carrier coordinate system. The extrinsic parameters include: the rotation matrix and translation vector of the 3D laser scanner to the carrier coordinate system; the intrinsic parameter matrix of the camera; and the rotation matrix and translation vector of the camera to the carrier coordinate system.

[0025] Through the above external parameter calibration, the spatial reference of the 3D laser scanner, high-definition camera and POS is unified, laying the foundation for subsequent spatiotemporal synchronous calculation.

[0026] S102, during train operation, uses the second pulse signal of the Global Navigation Satellite System as a reference to synchronously collect laser point cloud, image and positioning and attitude data along the track; The synchronous acquisition of laser point cloud, image, and positioning and attitude data along the track includes: using the second pulse signal of the global navigation satellite system as a reference, generating a synchronous trigger signal through frequency division by a field-programmable gate array, controlling the three-dimensional laser scanner and camera to acquire data synchronously at a preset frequency, so that the data from each sensor have a unified timestamp and the time error is less than a preset duration.

[0027] Specifically, in this embodiment, the carrier travels at a constant operating speed along the railway track, while a 3D laser scanner and a high-definition camera simultaneously acquire spatial geometric data and texture image data. During the acquisition process, it is ensured that the 3D laser scanner and the high-definition motion camera can acquire data simultaneously. A unified timestamp is generated using the Global Navigation Satellite System (GNSS) pulse-per-second (PPS) in the positioning and attitude determination system as the core reference, ensuring strict alignment of the data from each sensor in the time dimension. Within the same time dimension, the 3D laser scanner acquires 3D point information, and the high-definition camera acquires video images.

[0028] The field-programmable gate array (FPGA) generates a TTL signal of a first preset frequency to trigger a 3D laser scanner, and simultaneously generates a square wave signal of a second preset frequency to trigger a high-definition camera to capture images. All data packets are written with timestamps synchronized based on a precise time protocol, ensuring that the data from the 3D laser scanner, high-definition camera, and positioning and attitude determination system have a unified timestamp and a time error of less than 1 millisecond.

[0029] Within the same time dimension, the 3D laser scanner collects 3D point information based on the time-of-flight ranging principle and generates 3D point coordinates in its own coordinate system; the high-definition camera continuously captures images along the track at a fixed sampling rate, realizing a spatiotemporal one-to-one correspondence between the lidar point cloud data and the camera image data.

[0030] During the data acquisition process, the raw data generated by the sensors is written to the vehicle's solid-state drive in real time as a binary stream. The 3D laser scanner generates binary point cloud files at preset time intervals, and each frame of the high-definition camera is saved as an image file and stored in folders according to time. Positioning and attitude determination data are stored in binary files in real time. Simultaneously, through the wireless communication module, the system transmits low-resolution preview images and equipment status information back to the ground monitoring center at preset intervals. The monitoring personnel confirm that the data quality is up to standard and that there are no dropped frames or sensor malfunction alarms.

[0031] The 3D laser scanner generates 3D point coordinates in its own coordinate system based on the time-of-flight ranging principle.

[0032] Where r is the laser ranging value. θ The horizontal scanning angle, This is the vertical scanning angle.

[0033] High-definition cameras continuously capture images along the track at a fixed sampling rate, achieving a spatiotemporal one-to-one correspondence between lidar point cloud data and camera image data.

[0034] S103, perform combined navigation calculation on the positioning and attitude data to obtain the train's position and attitude information at continuous moments, and transform the laser point cloud to the world coordinate system based on the extrinsic parameters.

[0035] Furthermore, the combined navigation calculation of the positioning and attitude data includes: using tightly coupled Kalman filtering to fuse global navigation satellite system data, inertial measurement unit data, and odometer pulse data to calculate the position, speed, and attitude angle of the train at continuous time intervals.

[0036] Specifically, after data acquisition is complete, the vehicle-mounted solid-state drive is connected to the ground processing server via a high-speed data interface. All data recorded during the acquisition process, including GPS data, inertial measurement unit (IMU) data, odometer pulse data, laser point cloud files, and image files, are imported into the server's storage area. The server reads the raw data files output by the positioning and attitude determination system, extracts GPS observations, IMU angular velocity and acceleration values, and odometer pulse counts, in preparation for post-processing calculations.

[0037] The positioning and attitude determination data are used for combined navigation calculations. The server employs a tightly coupled Kalman filter to fuse Global Navigation Satellite System (GNSS) data, Inertial Measurement Unit (INS) data, and odometer pulse data to calculate the train's position, velocity, and attitude angles at continuous time points. Specifically, a tightly coupled bidirectional smoothed Kalman filter algorithm is used to post-process the GNSS data, INS data, and odometer pulse data. GNSS data provides the absolute position and velocity observations of the vehicle at a preset sampling rate; INS data provides the vehicle's three-axis angular velocity and specific force information at a preset sampling rate; and odometer pulse data provides the vehicle's displacement increment along the track direction. The filtering algorithm jointly estimates the pseudorange and carrier phase observations of the Global Navigation Satellite System, the integral results of the inertial measurement unit, and the displacement constraints of the odometer. Through iterative processing of forward filtering and backward smoothing, it corrects the cumulative drift of the inertial calculation, compensates for the accuracy loss when the Global Navigation Satellite System signal is blocked, and generates the train's accurate three-dimensional position coordinates, three-dimensional velocity components, roll angle, pitch angle, and yaw angle at each second pulse moment. The position accuracy is better than the preset threshold.

[0038] The laser point cloud is transformed to the world coordinate system based on the extrinsic parameters. The point cloud processing program reads the laser point cloud file. For each laser point, it first retrieves the carrier pose information corresponding to that moment based on its timestamp. Using the rotation matrix and translation vector from the calibrated 3D laser scanner to the carrier coordinate system, the laser point is transformed from the scanner coordinate system to the carrier coordinate system. Subsequently, using the calculated rotation matrix and position vector from the carrier coordinate system to the world coordinate system, the laser point is transformed from the carrier coordinate system to the unified world coordinate system. The world coordinate transformation formula for the laser point cloud is:

[0039] Among them, P WR represents the final coordinates of a point in the world / engineering coordinate system. B W (t i ) for t i The rotation matrix P from the time-space coordinate system to the world coordinate system. B W (t i ) for t i The position of the time carrier in the world coordinate system.

[0040] Through the above point-by-point coordinate transformation, the coordinate unification and initial processing of all laser point clouds are completed, generating unified point cloud data for the entire line.

[0041] S104. Based on the pre-constructed standard track contour three-dimensional pipeline region, the converted laser point cloud is subjected to spatial constraint filtering to remove noise points located outside the three-dimensional pipeline region, thus obtaining an effective point cloud.

[0042] Specifically, to quickly remove noisy point clouds and improve 3D modeling efficiency, this embodiment proposes a rapid noise reduction method for point clouds based on spatial constraint filtering of standard track contours. This method utilizes the standard prior knowledge of railway track geometry, using a pre-constructed 3D pipe region as a spatial envelope constraint to quickly distinguish between internal and external noise points in the entire line's point cloud. This directly removes a large number of noise points such as adjacent trains, trees, and birds during the filtering stage, while preserving core structural features such as the rail top surface and rail web, providing a high signal-to-noise ratio data foundation for subsequent true-color fusion and model reconstruction.

[0043] Furthermore, the converted laser point cloud is subjected to spatial constraint filtering, such as... Figure 2 As shown, it includes the following steps: S2001. During periods without train operation, a static scanner is used to acquire standard track point cloud samples. Rail cross-sections are extracted along the longitudinal direction of the track at preset intervals. A random sampling consensus algorithm is used to fit the standard rail profile and expand it by a preset distance to generate a two-dimensional profile. Specifically, typical sections of the line are selected, such as the straight section at K180+000 and the curved section with a radius of 800 meters at K210+000. During periods without train operation, a FARO S350 high-precision static scanner is used for detailed scanning to obtain standard track point cloud samples with a resolution of 1.5 mm. Rail cross-sections are extracted along the longitudinal direction of the track at 0.5-meter intervals. The RANSAC random sampling consensus algorithm is used to fit a standard profile of 60 kg / m rail. This fitted profile is then expanded outward by 5-10 mm to form the core rail profile, and further expanded towards the sleeper edge to form a 2.8-meter wide track structure profile. If necessary, this can be further expanded to the track bed slope to form a full-section profile of the track bed, thereby generating two-dimensional profile templates suitable for different levels of filtering requirements.

[0044] S2002, Based on the line design parameters, the two-dimensional contour is spatially transformed and interpolated along the longitudinal direction of the track with a preset step size to generate a continuous three-dimensional pipeline region. Specifically, the generated two-dimensional contour, along with the horizontal curves, vertical curves, and superelevation data from the route design parameters, is input into the system. The system generates a continuous three-dimensional pipeline region along the target modeling mileage range, for example, from K150+000 to K250+000, using linear interpolation with a step size of 5 meters. The contour shape is used for straight sections, while for curved sections, spatial transformations are performed based on the curve radius and superelevation value. The contour cross-sections are rotated and translated to ensure a smooth transition between adjacent step sizes, forming a continuous spatial envelope that conforms to the actual route shape, and is stored in the standard contour library of the spatial constraint filtering module.

[0045] S2003, iteratively register the current measured point cloud with the predefined contour of the corresponding mileage, calculate the translation and rotation, and update the contour parameters if the median of the registration error is greater than the preset threshold. Before each formal filtering process, the program automatically extracts a small segment from the current measured point cloud, such as the section from K180+000 to K180+100, and performs iterative nearest-point coarse registration with the predefined contour of the corresponding mileage in the standard contour library. The calculated global translation Δ is (0.012 m, -0.008 m, 0.005 m) and rotation θ is 0.15 degrees, with a median registration error of 8 mm. Since 8 mm is less than the preset threshold of 15 mm, the system determines that the contour match is good and no contour parameter update is required; if the median registration error is greater than 15 mm, an adaptive update of the contour parameters is triggered, correcting the actual contour parameters of the current measured section to the standard contour library to ensure that the 3D pipeline region is consistent with the actual state of the current track.

[0046] S2004: Based on the mileage coordinates of each point, retrieve the corresponding contour section, determine whether the registered and compensated points are located within the three-dimensional pipeline area, remove points outside the area, retain points within the area and points less than a preset distance from the boundary, and perform voxel downsampling on the retained valid point cloud with a preset voxel side length.

[0047] In practice, the CUDA architecture is used to process each point in parallel across thousands of GPU cores: First, based on the point's mileage coordinates, such as K190+520.35, the corresponding 3D pipe cross-section at mileage K190+520 in the standard contour library is retrieved. This cross-section is the interpolation result of the straight segment contour and the 800-meter radius curved segment contour. Then, iterative nearest-point registration compensation is applied to the 3D pipe cross-section, i.e., an inverse transformation opposite to the global translation Δ and rotation θ is applied to the contour. Subsequently, it is determined whether the plane coordinates of the compensated point fall within the boundary of the 3D pipe cross-section. Finally, a point cloud cluster (approximately 1.2 million points) from a passing train on an adjacent line, a point cloud of distant trees (approximately 8 million points), and several flocks of birds (approximately 15,000 points) are all located outside the 3D pipe region and are judged as noise and immediately removed. Points on the core contours such as the top surface and web of the rail, as well as points within 2 millimeters of the contour boundary, are retained and marked. After filtering, the number of point clouds dropped sharply from 18 billion to 2.9 billion, with a rejection rate of over 99%, and the entire filtering process took about 0.8 seconds. Subsequently, the remaining 2.9 billion point clouds were downsampled using voxels with a side length of 5 cm, outputting a high signal-to-noise ratio point cloud with uniform density (approximately 40 million points), which served as input data for subsequent true-color fusion and model reconstruction.

[0048] S105, based on the pose information and external parameters, the effective point cloud is projected onto the image plane corresponding to the timestamp, pixel color values ​​are extracted and multi-frame weighted fusion is performed to generate a true color point cloud; Furthermore, before projecting the 3D points in the camera coordinate system onto the pixel plane, an octree spatial index is constructed, and the candidate image frame corresponding to each valid point is quickly located based on the octree.

[0049] Specifically, the image identifiers and corresponding camera pose parameters of all valid image frames are input into the system to construct an octree spatial index of depth 8, enabling rapid retrieval of candidate image frames near each point. This octree, with its root node in three-dimensional space, recursively divides the space into eight sub-cubes along the X, Y, and Z dimensions. Each leaf node records the image frame identifiers falling within that spatial range and their corresponding camera pose information. Through the octree spatial index, the search complexity is reduced from... O (N points ·N frames Reduced to O (N points ·log(N) frames This enables real-time colorization at the level of millions of points per second. For each point in a high signal-to-noise ratio point cloud, an octree is used to quickly search for the multiple image frames that are closest to its spatial location as candidate image frames.

[0050] Furthermore, the step of projecting the effective point cloud onto the image plane corresponding to the timestamp includes: transforming the effective point cloud from the world coordinate system to the camera coordinate system based on the extrinsic parameters and pose information; projecting the three-dimensional points in the camera coordinate system onto the pixel plane using the camera's intrinsic parameter matrix to obtain the corresponding pixel coordinates; and extracting the RGB color value of the pixel if the pixel coordinates are within the image range.

[0051] Specifically, for each valid point, based on the precise timestamp, the image frame and pose are matched, and using the extrinsic parameters obtained from calibration and the pose information calculated by integrated navigation, the point is transformed from the world coordinate system to the camera coordinate system. The formula for transforming the point cloud from the world coordinate system to the camera coordinate system is:

[0052] Where P C =(X C ,Y C Z C ) T Let R be the three-dimensional coordinates of the point in the camera coordinate system. C B R is the rotation matrix from the carrier coordinate system to the camera coordinate system. B W This is the rotation matrix from the world coordinate system to the carrier coordinate system.

[0053] Then, the camera coordinates are projected onto the pixel plane using the camera's intrinsic parameters:

[0054] Where f x f y c is the camera focal length parameter. x， c y Principal point offset parameter, Z C This represents the depth value of the point in the camera coordinate system.

[0055] If the pixel coordinates are within the image range, the RGB color value of that pixel is extracted. Specifically, it is determined whether the pixel coordinates (u, v) obtained from the projection are within the valid range of the image frame. If the pixel coordinates are within the image range, the RGB color value of that pixel position is directly extracted. R i , G i , B i The color value of a given point in space is used as a candidate color value; if the pixel coordinates are outside the image range, the color information of the candidate frame is discarded.

[0056] Furthermore, the multi-frame weighted fusion includes: for the same physical point covered by multiple frames of images, calculating the distance weight based on the depth value of the point in the camera coordinate system, and performing a weighted average of the RGB color values ​​extracted from each frame according to the distance weight.

[0057] Specifically, the same physical point may be covered by multiple image frames. To avoid color difference at the seams, a distance-weighted fusion strategy is adopted. For multiple candidate image frames found through octree search, the depth value of the point in the camera coordinate system of each candidate frame is calculated, and the distance weight is calculated accordingly. ; The merged RGB values ​​are: .

[0058] S106. Based on the true-color point cloud, triangulation is performed, and after mesh optimization, hole repair and texture mapping, a real-world 3D model of the railway track is output.

[0059] Specifically, in this embodiment, the true-color point cloud file generated in the aforementioned steps is imported into the model reconstruction module, and the automatic reconstruction process from point cloud to real-world 3D model is initiated, providing a measurable and interactive 3D data foundation for the digital management of railway maintenance.

[0060] Furthermore, the triangulation based on true-color point clouds includes: generating an initial mesh using a parallel Delaunay triangulation algorithm and setting a maximum triangle side length threshold; detecting boundary edges and filling holes using the minimum angle method, and simplifying the mesh using a quadratic error metric algorithm; normalizing the pixel coordinates of the mesh vertices to texture coordinates based on the width and height of the texture image, generating a textured real-world 3D model and outputting it in a preset format, while simultaneously generating a metadata file containing the coordinate system, accuracy, and mileage range.

[0061] Specifically, the model reconstruction module calls the parallel Delaunay triangulation algorithm in the CGAL library to read the coordinates and color information of all spatial points in the true-color point cloud. It sets the maximum triangle side length threshold to 0.2 meters and enables surface normal consistency constraints, performing parallel triangulation of the point cloud region by region to generate an initial irregular triangular mesh. This initial mesh fully preserves the geometric details of the track surface, with densely distributed triangular faces, accurately representing the undulations of the rail top surface, sleeper gaps, and ballast surface.

[0062] The initial TIN mesh generated after subdivision contains a large number of triangular faces. The system automatically performs boundary edge detection to identify non-penetrating small holes caused by sleeper gaps, local point cloud defects, or occlusion, such as hole areas with a diameter of less than 5 cm. For the detected hole boundary edges, the minimum angle method is used for filling to ensure the topological continuity of the model surface. At the same time, in order to reduce the model data volume and improve the efficiency of subsequent loading and rendering, a quadratic error metric algorithm is used to simplify the initial mesh. While preserving the contour features, the number of triangular faces is compressed to a preset proportion of the original number to generate a lightweight optimized mesh.

[0063] The optimized mesh vertices are normalized by mapping their texture coordinates, converting the pixel coordinates (u,v) of each vertex into normalized texture coordinates based on the width W and height H of its source image. The mapping formula is:

[0064] Where W and H are the width and height of the texture image, respectively, and (u, v) are the projected pixel coordinates, (u... t v t ) represents the normalized texture coordinates.

[0065] The system packages all mesh vertices with the newly generated texture coordinates to create a textured, realistic 3D model. It supports multiple industry-standard output formats, including OSGB, 3D Tiles, and OBJ, suitable for detailed viewing in ContextCapture, publishing on WebGIS platforms, and importing into general 3D software, respectively. Simultaneously, the system automatically generates model metadata files, recording coordinate system, plane mean square error, elevation mean square error, texture resolution, modeling mileage range, and processing log information, forming a complete model delivery document.

[0066] Secondly, this embodiment also proposes a rapid 3D modeling system for railway tracks based on spatial constraint filtering. Using data acquisition as a foundation, spatial constraint filtering as a tool, and true-color point cloud fusion as support, it constructs a system for rapid and complete output from the physical existence of railway tracks to a 3D reality model. This solves the problems of low model output efficiency, long cycle time, slow point cloud noise reduction, and low model fidelity in rapid track modeling. The system mainly includes a data acquisition subsystem, a preprocessing subsystem, and a model generation subsystem. Its functional framework is as follows: Figure 3 As shown, the system operation process is as follows: Figure 4 As shown.

[0067] 1. Motion sensing module This module consists of a 3D laser scanner (such as a 32-line or 64-line LiDAR), a high-definition motion camera (supporting 4K resolution and global shutter), and a positioning and attitude determination system (POS, including a GNSS receiver, fiber optic gyroscope IMU, and high-precision odometer), rigidly integrated and mounted on the bottom of an operating train or a dedicated inspection vehicle platform. Precision mechanical fixtures ensure the relative positions of each sensor are fixed and coordinate system calibration is completed (scanner → carrier, camera → carrier). Using the GNSS PPS pulse as a reference, a synchronization trigger signal is generated through FPGA frequency division, achieving time synchronization of data acquisition from the laser scanner, camera, and odometer with an error of less than 1ms. This module is responsible for synchronously acquiring raw point cloud, image, and pose data as the train moves.

[0068] 2. Data transmission module Raw data (binary format) is temporarily stored in a streaming manner on an onboard solid-state drive (capacity ≥ 2TB, write speed ≥ 500MB / s) to prevent data loss due to network interruptions. Simultaneously, keyframes (such as a low-resolution preview image per kilometer and device status logs) are transmitted back to the ground center in real time via a 4G / 5G communication module for monitoring data acquisition quality. After acquisition, the complete data is exported to the ground processing server via USB 3.0 or 10 Gigabit Ethernet.

[0069] 3. Data storage module A hierarchical storage strategy is adopted: raw data is stored in folders according to timestamps (laser point clouds are stored as .las / .bin, images as .jpg / .png, and POS data as .txt / .csv). Processed intermediate data (denoised point clouds and true-color point clouds) and the final models (OSGB / 3D Tiles / OBJ) are stored in a solid-state cache and a mechanical hard drive archive, respectively. A metadata database is established to record acquisition time, route mileage, sensor parameters, processing logs, etc., supporting subsequent retrieval and reproduction.

[0070] 4. Point cloud noise reduction module This invention implements the spatial constraint filtering method based on standard track profiles. The module internally includes: Standard profile library: Pre-stores three-dimensional pipeline regions M with different track types (60kg / m, 75kg / m, etc.) and different line parameters (straight line, curve radius, superelevation).

[0071] Dynamic registration unit: Before each modeling, a 100m measured point cloud is automatically selected and coarsely registered with the corresponding mileage contour using ICP, and the translation Δ and rotation θ are calculated; if the median registration error is >15mm, the contour parameters are automatically updated.

[0072] Contour filtering unit: processes each point P in parallel. WBased on its mileage coordinates, the corresponding profile section M(m) is retrieved. It is determined whether it is within the registered and compensated M(m). If it is, it is retained; otherwise, it is discarded. Points near the boundary (distance ≤ 2mm) are retained and marked.

[0073] Downsampling unit: performs voxel downsampling on the retained point cloud (voxel side length 5cm) and outputs a high signal-to-noise ratio point cloud (uniform density, noise removal rate >99%).

[0074] This module runs on a GPU parallel architecture and takes less than 0.5 seconds to process one million points.

[0075] 5. True-color point cloud fusion module Implement a fast point cloud image coloring method. The module includes: Projection mapping unit: Based on the precise timestamp, match the image frame and POS pose, transform each 3D point P_W to the camera coordinate system and project it onto the pixel plane.

[0076] Color extraction unit: directly extracts the RGB values ​​of the corresponding pixels; for overlapping areas of multiple frames, distance-weighted fusion is used to eliminate color differences at the seams.

[0077] Octree Index Unit: Constructs an octree spatial index to quickly locate the candidate image frame corresponding to each point, achieving real-time colorization at the level of millions of points per second.

[0078] Output full-domain true-color 3D point cloud (format .las or .ply).

[0079] 6. Model Reconstruction Module Based on true-color point clouds, an irregular triangular network (TIN) model is constructed using the Delaunay triangulation algorithm. Specifically, this includes: Mesh generation: Use the C++ CGAL or PCL library for fast triangulation, and set the maximum triangle side length (e.g., 0.2m) to avoid redundant meshes.

[0080] Hole repair: Holes are filled using boundary edge detection and the minimum angle method to ensure model integrity.

[0081] Texture mapping optimization: Normalize the texture coordinates of the mesh vertices, convert the corresponding pixel coordinates (u,v) of the vertex to (ut,vt), and generate textured OBJ or 3D Tiles models.

[0082] Output measurable and interactive 3D models of real-world scenes.

[0083] 7. Model Output Module Supports multiple industry-standard output formats: OSGB (for ContextCapture), 3DTiles (for WebGIS / Cesium), and OBJ (for general 3D software). It also provides model simplification options (such as retaining 90% of feature points) for lightweight display on mobile devices. Automatic model metadata files (including coordinate system, accuracy, mileage range, etc.) are generated during output.

[0084] Furthermore, to enable those skilled in the art to better understand and implement the present invention, a specific application example in a section (100 km long) of a heavy-haul freight railway line from K150+000 to K250+000 is provided below to illustrate the technical solution of the present invention. This example is intended to explain, rather than limit, the scope of protection of the present invention.

[0085] The prerequisites for the example are as follows: (1) The working line: a railway down line, from K150+000 to K250+000, with a total length of 100 kilometers. This section includes a variety of typical working conditions such as straight lines, curves (minimum curve radius 400m), and slopes (maximum gradient 12‰). The working vehicle: an SS4G type electric locomotive with the number "DZ-001" in normal operation pulling freight trains. (2) Environmental conditions: cloudy, good visibility, with some trees, fences and occasional passing trains on adjacent lines along the line.

[0086] Step 1: System Calibration and Initialization 1.1 Sensor Rigid Integration and Coordinate System Definition Behind the bottom bogie of the "DZ-001" locomotive, it is secured sequentially using specially designed triangular truss-type rigid clamps: 3D laser scanner: Livox Avia (equivalent 32 lines, field of view 70.4°×77.2°, ranging accuracy ±2cm) was selected. Its origin O was defined. L As the laser emission center, X L The axis points to the right in the direction of locomotive travel, Y L Axially upward, Z L Axially forward. High-definition action camera: Basler acA2440-35gm (5 megapixels, global shutter, 25fps), 12mm focal length lens. Its origin O is defined. C For the lens optical center, X C Axis to the right, Y C Downward along the axis, Z C Axis forward (camera coordinate system).

[0087] Positioning and Attitude Determination System (POS): Applanix APX-20 (supports GNSS (GPS + BeiDou) + fiber optic gyroscope IMU + odometer). Its origin O is defined.B For IMU measurement center, X B Axis to the right, Y B Axially upward, Z B Axial forward (carrier coordinate system).

[0088] 1.2 Coordinate system calibration (external parameter calibration) Calibration field operation: The calibration is carried out in a dedicated calibration field (with 20 infrared reflective targets with known precise coordinates (CGCS2000 coordinate system)) inside the locomotive depot.

[0089] Laser scanner calibration: Control the Livox Avia to scan all targets and obtain the target's position at the origin O. L The coordinates below. Simultaneously, the target's position at O ​​is measured using a high-precision total station. B The coordinates below. Solve for the rotation matrix R. L B Translation vector t L B .

[0090] Camera calibration: Images of the checkerboard calibration board in different poses within the calibration field are captured, and the camera intrinsic parameter matrix K is obtained using the Zhang Zhengyou calibration method. Simultaneously, the rotation matrix R from the camera to the carrier coordinate system is calculated using known target points. C B Translation vector t C B .

[0091] Result storage: Store the R calculated above. L B t L B R C B t C B The parameters are written into the system's configuration file l.

[0092] 1.3 Loading the Standard Track Profile Library Contour Generation: Two typical sections of this line, K180+000 (straight line) and K210+000 (curve with R=800m), were selected. During the track maintenance window, a FARO S350 high-precision static scanner was used to acquire standard track point clouds (1.5mm resolution). The rail cross-section (60kg / m) was extracted, and the core rail contour was expanded outward by 8mm. The contour was then expanded outward towards the edge of the sleepers to form a track structure contour with a width of 2.8m.

[0093] 3D pipeline construction: Input the above 2D contours (both straight lines and curves with R=800m) along with the route design parameters (horizontal curves, vertical curves, superelevation) into the system. The system generates a continuous 3D pipeline region M along the mileage from K150+000 to K250+000, with a step size of 5m, through linear interpolation, and stores it in the standard contour library of the spatial constraint filtering module.

[0094] 1.4 Startup and Self-Test Before the locomotive leaves the depot, the modeling system control software is started. The software loads configuration file 1 and the standard contour library, and automatically detects the communication status of the Livox Avia, Basler camera, and APX-20, all of which display "Ready". The system time is synchronized with the GNSS satellite time.

[0095] Step Two: Dynamic Data Acquisition 2.1 Data Acquisition Start: The locomotive departs from K150+000 and maintains a stable speed of 65-72 km / h. The operator clicks "Start Data Acquisition" on the software interface.

[0096] 2.2 Synchronization Trigger: The system uses the GNSS PPS (pulses per second) output by the APX-20 as a reference, and generates a 20Hz TTL signal after frequency division by the FPGA to trigger the Livox Avia scan; it also generates a 25Hz square wave signal to trigger the Basler camera to capture images. All data packets are written with microsecond-level timestamps based on PTP (Precise Time Protocol) synchronization (e.g., 2026-05-10 13:05:23.124567).

[0097] 2.3 Data Streaming: Raw data generated by the sensors is written to the in-vehicle Samsung PM1735 SSD in real time as a binary stream. The Livox Avia generates a .lvx file (approximately 300MB) every 10 seconds, containing approximately 2 million points. Each frame of the Basler camera is saved as a .jpg file (approximately 0.5MB / frame), generating 12.5MB of data per second at 25fps, stored in folders by minute (Cam_Data / 20240515_1305 / ). POS data (position, attitude, velocity, timestamp) is stored in real time in .apx binary files.

[0098] 2.4 Data Backhaul and Monitoring: Through the 5G CPE equipment on the locomotive, the system transmits a low-resolution (640x480) preview image and status information such as equipment temperature and storage capacity to the ground monitoring center every kilometer. The monitor confirms that the data quality is qualified and there are no frame drops or sensor malfunction alarms.

[0099] Step 3: Trajectory Calculation and Point Cloud Preprocessing 3.1 Data Import: After data collection, the vehicle-mounted SSD is connected to the ground processing server via a USB interface (10Gbps). Copying all raw data (approximately 3.2TB) from the 100km distance takes approximately 15 minutes.

[0100] 3.2 Integrated Navigation Calculation: The server employs a tightly coupled bidirectional smooth Kalman filter algorithm to post-process and calculate the GNSS (5Hz sampling rate), IMU (200Hz), and odometer (pulse) data in the .apx file. The calculation results generate the locomotive's precise 6-DOF pose (x, y, z, roll, pitch, yaw) at each PPS time, with a position accuracy better than ±3cm.

[0101] 3.3 Original Point Cloud Coordinate Transformation: The point cloud processing program on the server reads the .lvx file. For each point P... L According to its timestamp t i Retrieve the calculated pose matrix R B W (t i (World → Carrier) and P B W (t i ), and according to the formula Convert it to the CGCS2000 world coordinate system. Generate a temporary, noisy full-line point cloud file (.las format, approximately 18 billion points, storage approximately 150GB).

[0102] 3.4 Spatial Constraint Filtering (Core Denoising Step) 1. Dynamic Registration: The program automatically extracts the measured point cloud segment from K180+000 to K180+100 and performs ICP (Iterative Closest Point) coarse registration with the predefined contour of the corresponding mileage in the standard contour library. The global translation Δ = (0.012m, -0.008m, 0.005m) and rotation θ = 0.15 degrees are calculated. The median registration error is 8mm (less than the 15mm threshold), indicating that the contour match is good and no update is required.

[0103] 2. Parallel Filtering and Removal: Utilizing the CUDA architecture, the GPU's 4096 cores process each point P in parallel. W First, according to P W Given the mileage coordinates (e.g., K190+520.35), retrieve the corresponding 3D pipe section M(m) at the mileage (K190+520) in the standard profile library (this is the interpolation result of the straight segment profile and the curved segment profile with R=800m). Then, after ICP registration compensation (applying -Δ and -θ to the profile M(m), determine P. Wwhether the plane coordinates (x, y) fall within the boundary of M(m). The final output results: a point cloud cluster from passing trains on adjacent lines (about 1.2 million points), a distant tree point cloud (about 8 million points) and several groups of flying bird point clouds (about 15,000 points) are all located outside M(m), which are determined as noise and immediately removed. Points on core contours such as the top surface of the steel rail and the rail waist (points within ≤ 2mm from the contour boundary) are retained and marked. After filtering, the number of point clouds dropped sharply from 18 billion to 2.9 billion (removal rate exceeding 99%). The entire filtering process takes about 0.8 seconds.

[0104] 3.5 Full-line Point Cloud Generation For the retained 2.9 billion point clouds, downsampling is performed with a voxel side length of 5cm. The final high signal-to-noise ratio point cloud (in .las format, about 40 million points, storage of about 3.6GB) is output, and the full-line point cloud is generated.

[0105] Step 4: True-color Point Cloud Fusion 4.1 Data alignment: the true-color point cloud fusion module reads the high signal-to-noise ratio point cloud .las file and all .jpg image files. For each point cloud frame (with timestamp t i ), find the image frame closest in time.

[0106] 4.2 Octree spatial index: input the image ID of all valid image frames and the corresponding camera pose R B W (t i ) and P B W (t i ) to construct an octree index with a depth of 8, so as to quickly retrieve candidate image frames near each point.

[0107] 4.3 Projection coloring (multi-frame fusion): for each point P in the high signal-to-noise ratio point cloud W (CGCS2000 coordinates): 3 closest image frames (Frame1, Frame2, Frame3) are quickly searched through the octree. For each candidate frame, convert P W to the camera coordinate system in sequence , and then project to the pixel plane . If the projection point (u, v) is within the image range of the frame (e.g., 0<u<2592, 0<v<1944), extract the RGB values (R i , G i , B i ) of the pixel, and calculate the weight ω i . The final RGB value of this point is the weighted average: , the same calculation applies to G and B. Output result: for example, a point on the top surface of the steel rail, in Frame1 its ZC =5.2m, Z in Frame2 C =5.5m, close-up images provide higher weight for clear textures. The entire process takes about 5 minutes and generates a true-color point cloud (.ply format, about 40 million points, file size about 4.2GB) containing X, Y, Z, R, G, and B information.

[0108] Step 5: Reconstruction of the 3D TIN Model 5.1 Triangulation: Import the true-color point cloud .ply file into the model reconstruction module. The module calls the parallel Delaunay triangulation algorithm in the CGAL library. Parameter settings: maximum triangle side length 0.2m, surface normal consistency enabled.

[0109] 5.2 Mesh Optimization: The initial TIN mesh generated after subdivision (approximately 80 million triangular faces). The system automatically performs hole repair: using a boundary edge detection algorithm to identify non-penetrating small holes (diameter <5cm) caused by sleeper gaps, and filling them using the minimum angle method. Simultaneously, mesh simplification is performed: using the Quadratic Error Metric (QEM) algorithm, the number of triangular faces is compressed to 12 million (preserving contour features), generating a lightweight model.

[0110] 5.3 Texture Mapping Optimization: The pixel coordinates (u, v) of each TIN mesh vertex are mapped according to the W(2592) and H(1944) of its source image, using the formula u t =u / W, v t =v / H normalizes to texture coordinates. Packs all mesh vertices with the newly generated texture coordinates.

[0111] 5.4 Model Output: The system simultaneously generates the textured model in three formats: Dazhun_Railway_3D_Model.osgb (4.5GB), for detailed viewing in ContextCapture; Dazhun_Railway_3D_Model.3dtiles (3.2GB), for publishing to WebGIS platforms (such as Cesium); and Dazhun_Railway_3D_Model.obj (with accompanying .mtl and .jpg texture atlases, totaling 5.8GB), for importing into general 3D software. Simultaneously, Dazhun_Railway_Model_Report.txt is generated, containing a model accuracy report: planar mean square error ±4.2cm, elevation mean square error ±3.8cm, and texture resolution 0.05m / pixel.

[0112] Step Six: Model Delivery and Application 6.1 Platform Import: Engineers upload the generated .3dtiles files to the railway bureau's internal digital twin platform.

[0113] 6.2 Application Example: 1. Track geometry detection: At K220+350, the system automatically measured the track gauge at this point to be 1437.2mm (the standard is 1435mm), triggering an over-limit alarm.

[0114] 2. Encroachment analysis: At K195+200, the model showed that a tree branch was only 1.2m away from the contact wire, triggering a safety warning. The maintenance personnel then arranged for pruning.

[0115] 3. Asset Management: Clicking on each fastener and sleeper in the model will bring up information such as its model number, installation date, and last maintenance record.

[0116] From the initial data acquisition at K150+000 to the final generation of the full-line 3D model, the total time was only 6 hours and 20 minutes. The core processing (filtering, fusion, and modeling) took approximately 1 hour and 40 minutes. Compared to the traditional static scanning + manual stitching method (which requires 2-4 weeks and more than 3 track maintenance windows for the same length), this method improves efficiency by more than 50 times, and causes no disruption to normal railway operations throughout the entire process.

[0117] Secondly, this embodiment also proposes a 3D modeling device for railway tracks based on spatial constraint filtering, used to execute the method described in the first aspect, such as... Figure 5 As shown, it includes: The calibration module 501 is used to calibrate the external parameters of the 3D laser scanner, camera, and positioning and attitude determination system to the carrier coordinate system. The acquisition module 502 is used to synchronously acquire laser point cloud, image and positioning and attitude data along the track during train operation, based on the second pulse signal of the global navigation satellite system. The navigation module 503 is used to perform combined navigation calculations on the positioning and attitude data, obtain the position and attitude information of the train at continuous time intervals, and convert the laser point cloud to the world coordinate system based on the extrinsic parameters. The filtering module 504 is used to perform spatial constraint filtering on the converted laser point cloud based on the pre-constructed standard track contour three-dimensional pipe region, and remove noise points located outside the three-dimensional pipe region to obtain an effective point cloud; The fusion module 505 is used to project the effective point cloud onto the image plane with the corresponding timestamp according to the pose information and external parameters, extract the pixel color values ​​and perform multi-frame weighted fusion to generate a true color point cloud. The reconstruction module 506 is used to perform triangulation based on the true-color point cloud, and after mesh optimization, hole repair and texture mapping, output a real-world 3D model of the railway track.

[0118] This device can be used to perform the 3D modeling method for railway tracks based on spatial constraint filtering as described in the first aspect, which will not be elaborated further here.

[0119] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for 3D modeling of railway tracks based on spatial constraint filtering, characterized in that, include: Calibrate the extrinsic parameters of the 3D laser scanner, camera, and positioning and attitude determination system to the carrier coordinate system; During train operation, laser point cloud, images, and positioning and attitude determination data along the track are collected synchronously based on the second pulse signal of the global navigation satellite system. The positioning and attitude data are combined for navigation calculation to obtain the train's position and attitude information at continuous moments, and the laser point cloud is converted to the world coordinate system based on the extrinsic parameters; Based on a pre-constructed standard track contour 3D pipeline region, spatial constraint filtering is applied to the converted laser point cloud to remove noise points located outside the 3D pipeline region, thus obtaining an effective point cloud; Based on the pose information and extrinsic parameters, the effective point cloud is projected onto the image plane corresponding to the timestamp, the pixel color values ​​are extracted and multi-frame weighted fusion is performed to generate a true color point cloud; Triangulation is performed based on true-color point clouds, and after mesh optimization, hole repair and texture mapping, a real-world 3D model of railway track is output. The step of performing spatial constraint filtering on the converted laser point cloud includes: During periods when no trains are running, a static scanner is used to acquire standard track point cloud samples. Rail cross-sections are extracted along the longitudinal direction of the track at preset intervals. A random sampling consensus algorithm is used to fit the standard rail profile and expand it by a preset distance to generate a two-dimensional profile. Based on the route design parameters, the two-dimensional contour is spatially transformed and interpolated along the longitudinal direction of the track with a preset step size to generate a continuous three-dimensional pipeline region. The current measured point cloud is iteratively registered with the predefined contour of the corresponding mileage, and the translation and rotation are calculated. If the median registration error is greater than the preset threshold, the contour parameters are updated. Based on the mileage coordinates of each point, the corresponding contour section is retrieved, and it is determined whether the registered and compensated points are located within the three-dimensional pipeline area. Points located outside the area are removed, and points located within the area and those less than a preset distance from the boundary are retained. The retained valid point cloud is then downsampled using voxel downsampling with a preset voxel side length. The extrinsic parameters include: the rotation matrix and translation vector of the 3D laser scanner to the carrier coordinate system; the intrinsic parameter matrix of the camera; and the rotation matrix and translation vector of the camera to the carrier coordinate system. The synchronous acquisition of laser point cloud, image, and positioning and attitude data along the track includes: using the second pulse signal of the global navigation satellite system as a reference, generating a synchronous trigger signal through frequency division by a field-programmable gate array, controlling the three-dimensional laser scanner and camera to acquire data synchronously at a preset frequency, so that the data from each sensor have a unified timestamp and the time error is less than a preset duration.

2. The method for 3D modeling of railway tracks based on spatial constraint filtering according to claim 1, characterized in that, The step of performing integrated navigation calculation on the positioning and attitude determination data includes: By employing tightly coupled Kalman filtering and fusing data from the Global Navigation Satellite System, the Inertial Measurement Unit, and the Odometer Pulse Data, the position, speed, and attitude angle of the train at continuous time points are calculated.

3. The method for 3D modeling of railway tracks based on spatial constraint filtering according to claim 1, characterized in that, The process of projecting the effective point cloud onto the image plane corresponding to the timestamp includes: Based on the extrinsic parameters and pose information, the effective point cloud is transformed from the world coordinate system to the camera coordinate system; Using the camera's intrinsic parameter matrix, three-dimensional points in the camera coordinate system are projected onto the pixel plane to obtain the corresponding pixel coordinates; If the pixel coordinates are within the image range, then the RGB color value of that pixel is extracted.

4. The method for 3D modeling of railway tracks based on spatial constraint filtering according to claim 3, characterized in that, Before projecting the 3D points in the camera coordinate system onto the pixel plane, an octree spatial index is constructed, and the candidate image frame corresponding to each valid point is quickly located based on the octree spatial index.

5. The method for 3D modeling of railway tracks based on spatial constraint filtering according to claim 4, characterized in that, The multi-frame weighted fusion includes: For the same physical point covered by multiple frames of images, a distance weight is calculated based on the depth value of the point in the camera coordinate system, and the RGB color values ​​extracted from each frame are weighted and averaged according to the distance weight.

6. The method for 3D modeling of railway tracks based on spatial constraint filtering according to claim 1, characterized in that, The triangulation based on true-color point clouds includes: The initial mesh is generated using a parallel Delaunay triangulation algorithm, and a threshold for the maximum triangle side length is set. Detect boundary edges and fill holes using the minimum angle method, and simplify the mesh using a quadratic error metric algorithm; The pixel coordinates of the mesh vertices are normalized to texture coordinates based on the width and height of the texture image, generating a textured real-world 3D model and outputting it in a preset format. At the same time, a metadata file containing the coordinate system, accuracy, and mileage range is generated.

7. A 3D modeling device for railway tracks based on spatial constraint filtering, used to execute the method described in any one of claims 1-6, characterized in that, include: The calibration module is used to calibrate the extrinsic parameters of the 3D laser scanner, camera, and positioning and attitude determination system to the carrier coordinate system; The acquisition module is used to synchronously acquire laser point cloud, imagery, and positioning and attitude data along the track during train operation, based on the second pulse signal of the Global Navigation Satellite System. The navigation module is used to perform combined navigation calculations on the positioning and attitude data, obtain the position and attitude information of the train at continuous time intervals, and convert the laser point cloud to the world coordinate system based on the extrinsic parameters. The filtering module is used to perform spatial constraint filtering on the converted laser point cloud based on a pre-constructed standard track contour three-dimensional pipeline region, to remove noise points located outside the three-dimensional pipeline region, and obtain an effective point cloud; The fusion module is used to project the effective point cloud onto the image plane with the corresponding timestamp based on the pose information and extrinsic parameters, extract pixel color values ​​and perform multi-frame weighted fusion to generate a true color point cloud; The reconstruction module is used to perform triangulation based on the true-color point cloud, and after mesh optimization, hole repair and texture mapping, output a real-world 3D model of the railway track.

Images