High-speed railway ballastless track construction measurement method
By combining flexible checkerboard calibration with IMU attitude data, the problem of large measurement errors in the construction of ballastless track for high-speed railways was solved, and high-precision track parameter measurement and construction acceptance data generation were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CCCC SECOND HIGHWAY ENG CO LTD
- Filing Date
- 2025-06-23
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies for track smoothness and gauge measurement in high-speed railway ballastless track construction suffer from problems such as complex camera parameter calibration, insufficient sensor data synchronization accuracy, and mismatch in weak texture areas, resulting in large measurement errors and failing to meet high-precision requirements.
The camera parameters are automatically calibrated using a flexible checkerboard calibration board. Spatiotemporal alignment is performed using IMU attitude data. A lightweight convolutional neural network is used for stereo matching to generate a high-precision orbit point cloud map. The track gauge and vertical displacement sequence are calculated to generate a track gauge table and a ride comfort table.
It achieves high-precision measurement of track parameters, eliminates cumulative errors in dynamic measurement, improves calibration efficiency and stereo matching accuracy, and the generated data can be directly used for construction acceptance standards.
Smart Images
Figure CN120608435B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data measurement technology, and more specifically, to a method for measuring the construction of ballastless track for high-speed railways. Background Technology
[0002] With the rapid development of high-speed railways, ballastless track has become the mainstream track structure for lines with speeds of 300 km / h and above due to its advantages of high smoothness, low maintenance costs and long-term stability. As the foundation for train operation, the smoothness and gauge of the track directly affect the stability and safety of train operation.
[0003] Currently, track smoothness and gauge measurements during the construction phase largely rely on track inspection trolleys. However, this method has several drawbacks: First, camera parameter calibration often employs traditional methods, which are complex and result in significant camera calibration errors. Second, insufficient sensor data synchronization leads to millisecond-level time discrepancies between binocular images, IMU attitude data, and GNSS positioning data, causing substantial error accumulation during dynamic measurements and hindering high-precision defect localization. Finally, binocular image processing often utilizes traditional stereo matching algorithms (such as SAD and SGM), which are prone to mismatches in areas with weak track texture (such as the rail surface or sides), resulting in sparse or distorted 3D point clouds and hindering accurate extraction of gauge and vertical height parameters. Therefore, this invention proposes a high-speed railway ballastless track construction measurement method to address these issues. Summary of the Invention
[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a method for surveying the construction of ballastless track for high-speed railways, comprising:
[0005] S1. Perform intrinsic and extrinsic parameter calibration on the binocular camera mounted on the track inspection trolley to obtain the corresponding intrinsic and extrinsic parameter data;
[0006] S2. The track detection vehicle acquires binocular images, mileage data, IMU attitude data and GNSS positioning data and performs spatiotemporal alignment processing.
[0007] S3. Construct a predicted state vector based on the spatiotemporally aligned mileage data, IMU attitude data, and GNSS positioning data;
[0008] S4. Perform stereo matching on the binocular images to generate a predicted disparity map; perform three-dimensional reconstruction on the predicted disparity map based on intrinsic data, extrinsic data and predicted state vector to generate an orbital point cloud map.
[0009] S5. Extract the coordinate sequence of the inner track points and the track elevation sequence from the track point cloud map; calculate the track gauge sequence based on the coordinate sequence of the inner track points, and generate the corresponding track gauge table based on the track gauge sequence;
[0010] S6. Perform vertical calculations on the IMU attitude data after spatiotemporal alignment to obtain the vertical displacement sequence of the track; perform weighted fusion of the track altitude sequence and the vertical displacement sequence to obtain the vertical height data of the track, and generate the corresponding smoothness table based on the vertical height data of the track; integrate the track gauge table and the smoothness table to generate a measurement report.
[0011] Furthermore, the methods for obtaining the intrinsic parameter data include:
[0012] The calibration image, which includes a flexible checkerboard calibration board and the track, is acquired by a track inspection trolley, and the IMU calibration data is recorded during the acquisition process. The coordinates of the checkerboard corner points in the flexible checkerboard calibration board in the calibration image are obtained, and the IMU rotation matrix is obtained by integrating the IMU calibration data.
[0013] Based on the coordinates of the chessboard corner points and the IMU rotation matrix, a projection error function is constructed. The intrinsic parameter matrix and distortion coefficients of the historical calibration are updated by minimizing the projection error function to obtain the intrinsic parameter matrix and distortion coefficients. The intrinsic parameter matrix and distortion coefficients together constitute the intrinsic parameter data.
[0014] Furthermore, the methods for obtaining the external parameter data include:
[0015] Obtain feature point pairs and rail gauge from the calibration image; construct an extrinsic parameter optimization function based on the feature point pairs and rail gauge; iteratively optimize the extrinsic parameter optimization function using the Levenberg-Marquardt method; when the extrinsic parameter optimization function iteratively converges to the minimum value, obtain the corresponding baseline, rotation matrix, and translation vector; the baseline, rotation matrix, and translation vector together constitute the extrinsic parameter data.
[0016] N sets of test images are acquired using a stereo camera with calibrated extrinsic parameters. Test matching point pairs in the test images are obtained through a feature matching algorithm. The vertical distance from the test matching point pair to the epipolar line is calculated and denoted as the epipolar residual. The average value of all epipolar residuals is calculated. When the average value is greater than or equal to the preset pixel threshold, the extrinsic parameters need to be recalibrated.
[0017] Furthermore, the methods for performing spatiotemporal alignment processing include:
[0018] The cable transmission delay and the time it takes for the image to be transmitted from the camera sensor to the processor are measured; the timestamp of the stereo camera exposure is corrected based on the cable transmission delay and the time it takes for the image to be transmitted from the camera sensor to the processor, and the corrected timestamp of the stereo camera exposure is obtained.
[0019] Based on the corrected timestamp of the stereo camera exposure, GNSS positioning data, IMU attitude data, and odometer data are aligned with the timestamp of the stereo image using interpolation.
[0020] Furthermore, the method for obtaining the predicted state vector includes:
[0021] The IMU attitude data and GNSS positioning data after spatiotemporal alignment are fused by a filtering algorithm to obtain the position vector, velocity vector, attitude quaternion vector and IMU sensor error vector of the track detection vehicle in the world coordinate system; the position vector, velocity vector, attitude quaternion vector and IMU sensor error vector are then concatenated to obtain the state vector.
[0022] A kinematic model is constructed, and the kinematic model is discretized to obtain a discrete-time prediction model. The state vector is used as the input to the discrete-time prediction model to obtain the motion state vector. Based on the spatiotemporally aligned mileage data, the position vector in the motion state vector is corrected by a filtering algorithm to obtain the predicted state vector.
[0023] Furthermore, the method for performing stereo matching includes:
[0024] The stereo image is calibrated based on the distortion coefficient in the intrinsic parameter data to obtain the calibrated stereo image; multi-scale feature extraction is performed on the calibrated stereo image through a lightweight convolutional neural network to obtain a multi-scale feature map;
[0025] Cost calculation is performed on the multi-scale feature map to obtain a hybrid cost volume; pixels in the hybrid cost volume are selected based on a winner-take-all strategy to obtain an initial disparity map; edge constraint optimization is performed on the initial disparity map using a conditional random field energy function and a parabolic interpolation function to obtain an optimized disparity map; the optimized disparity map is upsampled using a pre-built convolutional neural network to generate a predicted disparity map.
[0026] Furthermore, the method for obtaining the orbital point cloud map includes:
[0027] Based on the obtained intrinsic parameter matrix and baseline, triangulation calculation is performed on the predicted disparity map to obtain the coordinates of the pixels in the predicted disparity map in the camera coordinate system; the coordinates of the pixels in the camera coordinate system are transformed into coordinates in the IMU coordinate system through the obtained rotation matrix and translation vector; based on the attitude quaternion in the predicted state vector, the coordinates of the pixels in the IMU coordinate system are transformed into coordinates in the world coordinate system, and the coordinates of all pixels in the world coordinate system constitute a point cloud map; the point cloud map is filtered to obtain the orbit point cloud map.
[0028] Furthermore, the methods for obtaining the coordinate sequence of the inner points of the orbit and the orbital altitude sequence include:
[0029] Based on the track point cloud map, slices are made along the track extension direction at preset slicing intervals. Point cloud data at the top surface of the track in each slice is extracted, and the arithmetic mean of the height coordinates is calculated based on the point cloud data at the top surface of the track. The obtained arithmetic mean is arranged in spatial order to obtain the track height sequence.
[0030] Establish a reference coordinate system at the highest point of the top surface of the track in the track point cloud map. In the reference coordinate system, the X-axis is along the extension direction of the track, the Y-axis points to the inside of the track, and the Z-axis is vertically upward along the XY plane.
[0031] A feature extraction plane is constructed by moving down the Z-axis by a preset distance. A sampling window is established along the extension direction of the track at a preset slice interval. The point cloud coordinates inside the track in the feature extraction plane are obtained through the sampling window, and the coordinate sequence of the points inside the track is obtained.
[0032] Furthermore, the method for generating the corresponding track gauge table based on the track gauge sequence includes:
[0033] The roll angle is obtained by performing angle transformation on the attitude quaternion vector in the predicted state vector; a gauge calculation function is constructed based on the track design superelevation and roll angle, and the coordinate sequence of the inner points of the track is calculated using the gauge calculation function to obtain the gauge sequence; the gauge sequence is compared with the preset standard gauge to generate a corrected gauge sequence; and a corresponding gauge table is generated based on the gauge sequence, the corrected gauge sequence, and the position vector in the predicted state vector.
[0034] Furthermore, the method for obtaining the vertical displacement sequence of the orbit includes:
[0035] A vertical displacement kinematic model is constructed, and the vertical displacement kinematic model is discretized to obtain a vertical time displacement prediction model. The IMU attitude data after spatiotemporal alignment is used as the input of the vertical time displacement prediction model to obtain vertical displacement data. The vertical displacement data is aligned with the design elevation of the track at preset intervals to generate a vertical displacement sequence.
[0036] The technical effects and advantages of this invention are as follows:
[0037] This embodiment automatically calibrates camera parameters using a flexible checkerboard calibration board combined with the inherent characteristics of the track, avoiding the manual calibration field setup required in traditional calibration processes and improving calibration efficiency. By introducing IMU attitude data, errors caused by calibration board movement or deformation are avoided. By constructing a synchronization error compensation model and a cubic spline interpolation algorithm to correct timestamp deviations of different data, sub-millisecond spatiotemporal alignment of binocular images, IMU attitude data, GNSS positioning data, and mileage data is achieved, eliminating cumulative errors caused by time asynchrony in dynamic measurements. A lightweight MobileNetV3 network is used to extract multi-scale features, and a hybrid cost volume is generated by combining an adaptive weight fusion strategy. The edge details of the initial disparity map are optimized using conditional random fields and parabolic interpolation functions, significantly improving stereo matching accuracy in weak texture scenes. Sub-millimeter-level 3D point cloud reconstruction of the rail surface is achieved, providing a reliable data foundation for track gauge and elevation parameter calculations. The track gauge table and smoothness table generated on this basis can be directly connected to construction acceptance standards and support high-precision defect location. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of a high-speed railway ballastless track construction measurement method according to the present invention.
[0039] Figure 2 This is a schematic diagram of the steps for generating the orbital point cloud map in this invention. Detailed Implementation
[0040] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1
[0041] Please see Figure 1 As shown in this embodiment, a method for surveying the construction of ballastless track for high-speed railways includes:
[0042] S1. Perform intrinsic and extrinsic parameter calibration on the binocular camera mounted on the track inspection trolley to obtain the corresponding intrinsic and extrinsic parameter data;
[0043] S2. The track detection vehicle acquires binocular images, mileage data, IMU attitude data and GNSS positioning data and performs spatiotemporal alignment processing.
[0044] S3. Construct a predicted state vector based on the spatiotemporally aligned mileage data, IMU attitude data, and GNSS positioning data;
[0045] S4. Perform stereo matching on the binocular images to generate a predicted disparity map; perform three-dimensional reconstruction on the predicted disparity map based on intrinsic data, extrinsic data and predicted state vector to generate an orbital point cloud map.
[0046] S5. Extract the coordinate sequence of the inner track points and the track elevation sequence from the track point cloud map; calculate the track gauge sequence based on the coordinate sequence of the inner track points, and generate the corresponding track gauge table based on the track gauge sequence;
[0047] S6. Perform vertical calculations on the IMU attitude data after spatiotemporal alignment to obtain the vertical displacement sequence of the track; perform weighted fusion of the track altitude sequence and the vertical displacement sequence to obtain the vertical height data of the track, and generate the corresponding smoothness table based on the vertical height data of the track; integrate the track gauge table and the smoothness table to generate a measurement report.
[0048] Specifically, the methods for calibrating the intrinsic parameters of the binocular camera mounted on the track inspection vehicle and obtaining the corresponding intrinsic parameter data include:
[0049] Specifically, using standard track gauge and track height as geometric references, a straight track section with a length greater than a preset length (e.g., 10m) and a track gauge error less than a preset track gauge error (e.g., 0.1mm) is selected and marked as the calibration section. A flexible checkerboard calibration plate (with a grid size of 20mm) is laid longitudinally along the calibration section to ensure that the flexible checkerboard calibration plate presents multiple postures (pitch angle and yaw angle ±20°) within the field of view of the binocular camera. The flexible checkerboard calibration plate is fixed with a non-invasive clamp to ensure that the flexible checkerboard calibration plate is free of wrinkles or offset.
[0050] Keeping the flexible checkerboard calibration board within the field of view of the binocular camera, the track inspection trolley moves at a preset speed (e.g., 0.5 m / s) at a constant speed. The binocular camera simultaneously acquires images containing the flexible checkerboard calibration board and the left and right rails at 15 fps, which are recorded as calibration images (divided into left-eye and right-eye images). IMU calibration data is recorded during the acquisition process. The checkerboard corners of the flexible checkerboard calibration board in the calibration images are detected by the checkerboard corner detection function (indChessboardCorners), and the calibration images are optimized to subpixel accuracy (e.g., optimized to 0.01 pixel precision) by the subpixel corner optimization function (cornerSubPix).
[0051] A reprojection error function is constructed, and the historically calibrated intrinsic parameter matrix and distortion coefficients are updated by minimizing the projection error function to obtain the intrinsic parameter matrix K and distortion coefficients D; the intrinsic parameter matrix K and distortion coefficients D together constitute the intrinsic parameter data;
[0052] The expression for the reprojection error function is: ;in; This indicates the reprojection error term; The ordinal number of the corner point of the chessboard grid in the flexible chessboard grid calibration board; This represents the observed pixel coordinates of the i-th corner point of the flexible checkerboard calibration board in the calibration image; This represents the coordinates of the corner point of the i-th grid on the flexible chessboard calibration board. This is a projection function used to project... Mapped to the image plane; K is the camera's intrinsic parameter matrix (3×3, obtained from the previous calibration), containing the camera's focal length and principal point; D is the camera's distortion coefficients (obtained from the previous calibration), containing radial and tangential distortion; Indicates the IMU-visual pose alignment term; This indicates the time-varying angular velocity obtained by integrating the angular velocity from the IMU measurement data. The IMU rotation matrix (3×3); This represents the camera rotation matrix calculated based on the intrinsic parameter matrix, distortion coefficients, and checkerboard corner points. The time corresponding to the IMU measurement data; The weighting coefficient is used to balance the contributions of the two errors. The larger the value, the more it emphasizes the consistency between the IMU and the visual pose; the smaller the value, the more it depends on the accuracy of visual reprojection. The value ranges from 0.05 to 0.2.
[0053] Methods for obtaining extrinsic data by calibrating the extrinsic parameters of a stereo camera include:
[0054] Feature points in the calibration image are extracted using the SIFT algorithm to obtain feature point pairs; the gauge of the rails in the calibration image is obtained using the SGM algorithm.
[0055] An extrinsic parameter optimization function is constructed based on feature point pairs and rail gauge. The function expression of the extrinsic parameter optimization function is as follows: ;in, This indicates the track gauge error term, and its function is to force the calibration results to conform to the standard track gauge. This indicates the track gauge of the rails in the calibration image; This represents the epipolar residual term, which serves to ensure the geometric consistency of binocular matching; This represents the pixel coordinates of the i-th feature point in the left eye image of the calibration image; Let represent the pixel coordinates of the i-th feature point in the right eye image of the calibration image; i is the ordinal number of the feature point; F represents the fundamental matrix, used to describe the epipolar geometry of the stereo camera, calculated from the stereo camera's rotation matrix R (3×3, obtained from the previous calibration), translation vector t (obtained from the previous calibration), and camera intrinsic parameters, expressed as: ,in and Let be the intrinsic parameter matrices of the left and right cameras in a stereo camera system. It is the antisymmetric matrix of the translation vector; The weighting coefficient (ranging from 0.1 to 0.3) is used to balance gauge error and geometric consistency in the polar residual term. This is an IMU-visual alignment term, used to suppress attitude drift caused by vibration. Represents the IMU rotation matrix; Represents the camera rotation matrix; The weighting coefficients for the IMU-visual alignment term (ranging from 0.03 to 0.1) are used to balance the weights of sensor fusion.
[0056] It should be noted that the epipolar line is determined by the extrinsic parameters (relative position and attitude) of the binocular camera. It represents the straight line on the image plane of the right (left) camera that contains all possible matching points corresponding to a point in the image of the left (right) camera, reflecting the geometric constraint relationship between the two cameras. The accuracy of the extrinsic parameter calibration can be verified by calculating the perpendicular distance from the matching point to the epipolar line (epidural residual).
[0057] The extrinsic parameter optimization function is iteratively optimized using the Levenberg-Marquardt method. When the extrinsic parameter optimization function converges to the minimum value, the baseline B (which is an inherent parameter of the camera and needs to be checked for changes through calibration), rotation matrix R, and translation vector t of the corresponding stereo camera are obtained. The baseline B, rotation matrix R, and translation vector t together constitute the extrinsic parameter data.
[0058] N sets of test images are acquired using a stereo camera with calibrated extrinsic parameters. At least 100 test matching point pairs are obtained in each set of test images using feature matching algorithms (such as SIFT, ORB). The fundamental matrix is calculated using the calibrated extrinsic parameters; for each feature point in the left eye image of the test image, its corresponding epipolar equation in the right eye image is calculated, and the expression for the epipolar equation is: Where F represents the fundamental matrix; calculate the test matching point pair. to the poles The vertical distance is denoted as the epipolar residual. The average value of the epipolar residuals of all test matching point pairs is calculated. If the average value of the epipolar residuals is less than 0.25 pixels, it means that the extrinsic parameters meet the requirements. If the average value of the epipolar residuals is greater than or equal to 0.25 pixels, the extrinsic parameters need to be recalibrated or the feature matching accuracy needs to be checked.
[0059] The track inspection vehicle acquires binocular images, odometer data, IMU attitude data, and GNSS positioning data (position in the world coordinate system) using sensors carried on board. These sensors include a binocular camera, an odometer, an IMU, and GNSS. The binocular camera is a global shutter CMOS camera (such as FLIR BlackflyS) with a resolution of 4000×3000, a frame rate of 15fps, a focal length of 12mm, a baseline of 0.3m, and is equipped with an 850nm near-infrared illuminator. The IMU uses a fiber optic gyroscope (such as Hon...). The system uses a eywell HG4930 receiver with zero-bias stability ≤0.01° / h and a sampling rate of 200Hz. The GNSS uses a dual-frequency RTK receiver (such as Trimble R12), supports multiple constellations, and has a positioning accuracy of 10mm ± 1ppm. A binocular camera is mounted at the front of the vehicle, with its optical axis parallel to the track and an elevation angle of 15° to 45°. The IMU is fixed to the vehicle's center of gravity, and the GNSS antenna is placed on the roof of the vehicle. The IMU attitude data includes raw Euler angles, raw acceleration measurements, and raw angular velocity measurements. The binocular images include left and right eye images.
[0060] Specifically, the track inspection vehicle travels along the track to be measured, acquires mileage data, IMU attitude data and GNSS positioning data, and triggers the stereo camera to expose through the GNSS PPS (pulses per second) signal, thereby acquiring stereo images of the track;
[0061] A synchronization error compensation model is constructed to correct the exposure timestamp of the stereo camera. The expression of the synchronization error compensation model is as follows: ;in This indicates the corrected camera exposure timestamp (aligned with GNSS time, in milliseconds). This indicates the timestamp of the PPS signal received by GNSS (UTC time base, unit: milliseconds). It represents the cable transmission delay, which is the physical transmission time of the PPS signal from the GNSS antenna to the camera, calibrated by round-trip time (RTT) (unit: milliseconds). This indicates the time it takes for an image to be transmitted from the camera sensor to the processor (including exposure, readout, and transmission links; unit: milliseconds).
[0062] Based on the corrected timestamp of the stereo camera exposure, GNSS positioning data, IMU attitude data, and odometer data are aligned with the stereo image timestamp using interpolation methods (such as cubic spline interpolation).
[0063] The above process, through hardware-triggered compensation and software interpolation algorithms, achieves high-precision time alignment (sub-millisecond level) of multi-sensor data, providing a reliable timing reference for subsequent track geometry parameter calculation and defect detection.
[0064] The methods for obtaining the predicted state vector include:
[0065] The IMU attitude data and GNSS positioning data, after being processed by spatiotemporal alignment, are fused using extended Kalman filtering or error-state Kalman filtering to output the position vector of the track detection vehicle in the world coordinate system. (Unit: meters) and velocity vector (Unit: m / s), attitude quaternion vector (This represents the rotational transformation from the IMU body coordinate system to the world coordinate system, obtained through conversion of the original Euler angle data) and the errors of the IMU sensors, including the accelerometer zero bias vector. (Represents the inherent hardware error of the accelerometer, unit: m / s²) and the gyroscope zero bias vector (This represents the low-frequency random error of the gyroscope, in radians per second).
[0066] The position vector, velocity vector, attitude quaternion vector, accelerometer zero-bias vector, and gyroscope zero-bias vector are concatenated to obtain the state vector. ;
[0067] Construct a kinematic model, the functional expression of which is: ;in This represents a vector composed of measured values of acceleration and angular velocity. This indicates the rate of change of the position of the track detection trolley, in meters per second. This represents the rate of change of the speed of the track inspection trolley, in meters per second². Represents the expression of attitude quaternion vectors The exported rotation matrix is used to adjust the IMU acceleration. Transform to world coordinate system; This represents the acceleration measurement in the spatiotemporally aligned IMU attitude data, in meters per second². Represents the gravitational acceleration vector ( (Unit: meters per second²) The rate of change of attitude of the track inspection vehicle is expressed in radians per second. Represents the original angular velocity measurement in the spatiotemporally aligned IMU attitude data, in radians per second; Represents quaternion multiplication; This indicates the rate of change of zero bias of the accelerometer, in meters per second³. This represents the rate of change of zero bias of the gyroscope, in radians per second². Indicates the zero-bias noise of the accelerometer, unit: meters per second³; This indicates the zero-bias noise of the gyroscope, in radians per second².
[0068] The dynamic model is transformed into a discrete-time prediction model through numerical integration (4th-order Runge-Kutta method), thereby obtaining the motion state vector; the expression of the discrete-time prediction model is: ;in Represents the discrete-time predicted state vector; This represents the state vector at the current time point; This represents a vector consisting of the measured acceleration and angular velocity values at the current moment. Indicates a point in time; This represents the equivalent state increment function calculated using RK4; The noise vector representing the discretization process follows a zero-mean Gaussian distribution. , It represents the noise covariance matrix composed of the position vector, velocity vector, attitude vector, accelerometer zero bias vector, and gyroscope zero bias vector, and is used to describe the noise statistical characteristics of each component in state prediction; ; This indicates the discretization time step (0.005 seconds), which is strictly synchronized with the IMU sampling rate;
[0069] Based on the spatiotemporally aligned mileage data, the position vector in the motion state vector is corrected using a filtering algorithm (such as Kalman filtering or mean filtering) to obtain the spatiotemporally aligned predicted state vector. .
[0070] like Figure 2 As shown, the steps for generating an orbital point cloud map include:
[0071] The binocular image is calibrated based on the distortion coefficient in the intrinsic parameters to obtain the calibrated binocular image.
[0072] Multi-scale feature extraction is performed on the calibrated stereo images using a lightweight convolutional neural network (MobileNetV3) to obtain multi-scale feature maps (with resolutions of 1 / 8, 1 / 4, and 1 / 2 of the stereo image).
[0073] Cost calculation is performed on the multi-scale feature map to obtain a hybrid cost volume; using a winner-take-all strategy, the disparity with the minimum cost value of each pixel is selected from the hybrid cost volume, which is used as the initial disparity estimation result, and finally the initial disparity map is obtained.
[0074] The initial disparity map is optimized by edge constraint using the conditional random field energy function (univariate term + binary term) and parabolic interpolation function to obtain an optimized disparity map that preserves the details of the rail profile.
[0075] The optimized disparity map is upsampled using a pre-built convolutional neural network to generate a predicted disparity map (with the same resolution as the binocular image).
[0076] Based on the obtained baseline and intrinsic parameter matrix, the triangulation calculation is performed on the predicted disparity map to obtain the coordinates of each pixel in the camera coordinate system in the predicted disparity map.
[0077] The rotation matrix and translation vector of the stereo camera are used to transform the coordinates of the pixels in the camera coordinate system into coordinates in the IMU coordinate system.
[0078] Based on the attitude quaternion vector in the predicted state vector, the coordinates of pixels in the IMU coordinate system are transformed into coordinates (X, Y, Z) in the world coordinate system. The coordinates of all pixels in the world coordinate system constitute a point cloud map.
[0079] The point cloud image is filtered to remove outliers (such as fasteners or other debris) and noise data (through statistical filtering, neighborhood standard deviation > 3σ) to obtain the track point cloud image.
[0080] It should be noted that the cost calculation methods include:
[0081] The initial matching cost is obtained by calculating the cross-channel correlation of multi-level feature maps using cosine similarity. Based on the initial matching cost, the local pixel intensity difference of the feature maps is supplemented by the traditional SAD algorithm (with a calculation window size of 5×5) to obtain the local cost.
[0082] The initial matching cost and local cost are weighted and fused using an adaptive weight fusion strategy to generate a hybrid cost volume. The hybrid cost volume consists of three three-dimensional cost matrices (height × width × disparity range) containing a disparity range of 0-256 pixels.
[0083] The camera coordinate system is defined as follows: Origin: the center of the baseline; X-axis: horizontally to the right along the imaging plane; Y-axis: vertically downward along the imaging plane; Z-axis: perpendicular to the camera imaging plane, pointing directly in front of the scene. The IMU coordinate system is defined as follows: Origin: the center of the sensor located in the inertial measurement unit (IMU); X-axis: horizontally to the right along the front of the sensor (or the direction of device movement); Y-axis: vertically upward along the front of the sensor; Z-axis: following the right-hand rule, perpendicular to the XY plane and pointing forward. Triangulation calculation and coordinate system transformation are processes well known to those skilled in the art, and the specific calculation process and coordinate transformation process will not be given here.
[0084] The methods for obtaining the coordinate sequence of points inside the orbit and the orbital altitude sequence include:
[0085] Based on the track point cloud map, slices are made along the track extension direction at preset slicing intervals (e.g., 0.01m). Point cloud data at the top surface of the track in each slice is extracted, and the arithmetic mean of the height coordinates is calculated based on the point cloud data at the top surface of the track. The obtained arithmetic mean values are arranged in spatial order to form a complete track height sequence.
[0086] A reference coordinate system is established based on the highest point group on the top surface of the track. In the reference coordinate system, the X-axis is along the extension direction of the track, the Y-axis points to the inside of the track, and the Z-axis is perpendicular to the XY plane and pointing upwards.
[0087] A feature extraction plane is constructed by moving it down the Z-axis by a preset distance (14mm or 16mm from the highest point of the track top surface). Sampling windows are established along the track extension direction at preset slice intervals. The point cloud coordinates of the inner side of the feature extraction plane are obtained through the sampling windows, thus obtaining the coordinate sequence of the inner points of the track. and ;
[0088] The roll angle is obtained by performing angle transformation on the attitude quaternion vector in the predicted state vector. Based on track design superelevation (derived from track design documents) and roll angle Construct a track gauge calculation function, and use this function to calculate the coordinate sequence of points inside the track to obtain the track gauge sequence. ; track gauge sequence Compare with the preset standard gauge (1435mm) to generate a corrected gauge sequence (how much it differs from the standard gauge), and combine it with the position vector in the predicted state vector to generate the corresponding gauge table (including position, gauge, and the corrected gauge required for that position).
[0089] It should be noted that the function expression for calculating track gauge is as follows: ;in and These are the coordinate sequences inside the left and right tracks, respectively. The track design is for ultra-high values.
[0090] Construct a vertical displacement kinematic model. The functional expression of the vertical displacement kinematic model is as follows: ;in This represents the derivative of the vertical (Z-axis) position of the track inspection trolley with respect to time (i.e., vertical velocity), in meters per second. This indicates the instantaneous vertical velocity of the track inspection trolley, measured in meters per second. This indicates the vertical acceleration of the track inspection trolley, measured in meters per second². Represents the vertical acceleration measurement in the spatiotemporally aligned IMU attitude data, in meters per second². This represents the vector of gravitational acceleration (in meters per second²). The pitch angle of the vehicle is calculated from the attitude quaternion vector in the predicted state vector, and the unit is radians. Accelerometer zero bias vector; This indicates the rate of change of zero bias of the accelerometer, in meters per second³. Indicates the zero-bias noise of the accelerometer, unit: meters per second³;
[0091] The vertical displacement kinematic model is transformed into a discrete vertical time displacement prediction model through numerical integration (4th-order Runge-Kutta method). The spatiotemporally aligned IMU attitude data is used as input to the vertical time displacement prediction model to obtain vertical displacement data. Every 1 km, the vertical displacement data obtained from the vertical time displacement prediction model is aligned with the track's design elevation to generate a vertical displacement sequence. ;
[0092] The track height sequence and vertical displacement sequence are weighted and fused (weight ratio of 0.45:0.55) to obtain the track vertical height data; a vertical curve is plotted based on the track vertical height data, where the horizontal axis of the vertical curve is the position corresponding to the vertical height data, and the vertical axis is the vertical height data.
[0093] Based on the track design documents, a smoothness standard curve is drawn and superimposed onto the vertical curve graph. First, an unevenness threshold is set. Then, from the drawn vertical curve graph, curve intervals where the absolute value of the difference between the actual curve and the smoothness standard curve exceeds the preset threshold are extracted. For each extracted curve interval, its interval position on the graph, the corresponding track vertical height data, and the absolute value of the difference between the actual curve and the standard curve are recorded in detail. Finally, a standardized track smoothness table is generated based on the above recorded information. The track gauge table and the smoothness table are integrated by position to generate a measurement report for ballastless track.
[0094] This embodiment automatically calibrates camera parameters using a flexible checkerboard calibration board combined with the inherent characteristics of the track, avoiding the manual calibration field setup required in traditional calibration processes and improving calibration efficiency. By introducing IMU attitude data, errors caused by calibration board movement or deformation are avoided. By constructing a synchronization error compensation model and a cubic spline interpolation algorithm to correct timestamp deviations of different data, sub-millisecond spatiotemporal alignment of binocular images, IMU attitude data, GNSS positioning data, and mileage data is achieved, eliminating cumulative errors caused by time asynchrony in dynamic measurements. A lightweight MobileNetV3 network is used to extract multi-scale features, and a hybrid cost volume is generated by combining an adaptive weight fusion strategy. The edge details of the initial disparity map are optimized using conditional random fields and parabolic interpolation functions, significantly improving stereo matching accuracy in weak texture scenes. Sub-millimeter-level 3D point cloud reconstruction of the rail surface is achieved, providing a reliable data foundation for track gauge and elevation parameter calculations. The track gauge table and smoothness table generated on this basis can be directly connected to construction acceptance standards and support high-precision defect location.
[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0096] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0097] In the description of this invention, it should be understood that the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.
[0098] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.
[0099] In the description of this invention, "several" means one or more, and "a large number" means two or more.
[0100] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0101] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.
[0102] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.
Claims
1. A method for surveying construction of ballastless track for high-speed railways, characterized in that, include: S1. Perform intrinsic and extrinsic parameter calibration on the binocular camera mounted on the track inspection trolley to obtain the corresponding intrinsic and extrinsic parameter data; The methods for obtaining the internal parameter data include: The calibration image containing the flexible checkerboard calibration board and the track is acquired by the track inspection trolley, and the IMU calibration data during the acquisition process is recorded; the coordinates of the checkerboard corner points in the flexible checkerboard calibration board in the calibration image are obtained, and the IMU rotation matrix is obtained by integrating the IMU calibration data; Based on the coordinates of the chessboard corner points and the IMU rotation matrix, a projection error function is constructed. The intrinsic parameter matrix and distortion coefficients of the historical calibration are updated by minimizing the projection error function to obtain the intrinsic parameter matrix and distortion coefficients. The intrinsic parameter matrix and distortion coefficients together constitute the intrinsic parameter data. The methods for obtaining the external parameter data include: Obtain feature point pairs and rail gauge from the calibration image; construct an extrinsic parameter optimization function based on the feature point pairs and rail gauge; iteratively optimize the extrinsic parameter optimization function using the Levenberg-Marquardt method; when the extrinsic parameter optimization function iteratively converges to the minimum value, obtain the corresponding baseline, rotation matrix, and translation vector; the baseline, rotation matrix, and translation vector together constitute the extrinsic parameter data. N sets of test images are acquired using a stereo camera with calibrated extrinsic parameters. Test matching point pairs in the test images are obtained through a feature matching algorithm. The vertical distance from the test matching point pair to the epipolar line is calculated and denoted as the epipolar residual. The average value of all epipolar residuals is calculated. When the average value is greater than or equal to the preset pixel threshold, the extrinsic parameters need to be recalibrated. S2. Acquire binocular images, odometer data, IMU attitude data, and GNSS positioning data using a track detection vehicle, and perform spatiotemporal alignment processing; the spatiotemporal alignment processing includes the following methods: The cable transmission delay and the time it takes for the image to be transmitted from the camera sensor to the processor are measured; the timestamp of the stereo camera exposure is corrected based on the cable transmission delay and the time it takes for the image to be transmitted from the camera sensor to the processor, and the corrected timestamp of the stereo camera exposure is obtained. Based on the corrected timestamp of the binocular camera exposure, GNSS positioning data, IMU attitude data, and odometer data are aligned with the timestamp of the binocular image using interpolation. S3. Construct a predicted state vector based on the spatiotemporally aligned mileage data, IMU attitude data, and GNSS positioning data; S4. Perform stereo matching on the binocular images to generate a predicted disparity map; reconstruct the predicted disparity map in three dimensions based on intrinsic parameter data, extrinsic parameter data, and the predicted state vector to generate an orbital point cloud map; the stereo matching method includes: The stereo image is calibrated based on the distortion coefficient in the intrinsic parameter data to obtain the calibrated stereo image; multi-scale feature extraction is performed on the calibrated stereo image through a lightweight convolutional neural network to obtain a multi-scale feature map; Cost calculation is performed on the multi-scale feature map to obtain a hybrid cost volume; pixels in the hybrid cost volume are selected based on a winner-take-all strategy to obtain an initial disparity map; edge constraint optimization is performed on the initial disparity map using a conditional random field energy function and a parabolic interpolation function to obtain an optimized disparity map; the optimized disparity map is upsampled using a pre-built convolutional neural network to generate a predicted disparity map. S5. Extract the coordinate sequence of the inner track points and the track elevation sequence from the track point cloud map; calculate the track gauge sequence based on the coordinate sequence of the inner track points, and generate the corresponding track gauge table based on the track gauge sequence; the method of generating the corresponding track gauge table based on the track gauge sequence includes: The roll angle is obtained by performing angle transformation on the attitude quaternion vector in the predicted state vector; a gauge calculation function is constructed based on the track design superelevation and roll angle, and the coordinate sequence of the inner points of the track is calculated using the gauge calculation function to obtain the gauge sequence; the gauge sequence is compared with the preset standard gauge to generate a corrected gauge sequence; a corresponding gauge table is generated based on the gauge sequence, the corrected gauge sequence and the position vector in the predicted state vector. S6. Perform vertical calculations on the spatiotemporally aligned IMU attitude data to obtain the vertical displacement sequence of the orbit; the method for obtaining the vertical displacement sequence of the orbit includes: A vertical displacement kinematic model is constructed, and the vertical displacement kinematic model is discretized to obtain a vertical time displacement prediction model. The IMU attitude data after spatiotemporal alignment is used as the input of the vertical time displacement prediction model to obtain vertical displacement data. The vertical displacement data is aligned with the design elevation of the track at preset intervals to generate a vertical displacement sequence. The track height sequence and vertical displacement sequence are weighted and fused to obtain the track vertical height data, and a corresponding smoothness table is generated based on the track vertical height data; the track gauge table and smoothness table are integrated to generate a measurement report.
2. The method for surveying construction of ballastless track for high-speed railways according to claim 1, characterized in that, The methods for obtaining the predicted state vector include: The IMU attitude data and GNSS positioning data after spatiotemporal alignment are fused by a filtering algorithm to obtain the position vector, velocity vector, attitude quaternion vector and IMU sensor error vector of the track detection vehicle in the world coordinate system; the position vector, velocity vector, attitude quaternion vector and IMU sensor error vector are then concatenated to obtain the state vector. A kinematic model is constructed, and the kinematic model is discretized to obtain a discrete-time prediction model. The state vector is used as the input to the discrete-time prediction model to obtain the motion state vector. Based on the spatiotemporally aligned mileage data, the position vector in the motion state vector is corrected by a filtering algorithm to obtain the predicted state vector.
3. The method for surveying construction of ballastless track for high-speed railway according to claim 2, characterized in that, The methods for obtaining the orbital point cloud map include: Based on the obtained intrinsic parameter matrix and baseline, triangulation calculation is performed on the predicted disparity map to obtain the coordinates of the pixels in the predicted disparity map in the camera coordinate system; the coordinates of the pixels in the camera coordinate system are transformed into coordinates in the IMU coordinate system through the obtained rotation matrix and translation vector; based on the attitude quaternion in the predicted state vector, the coordinates of the pixels in the IMU coordinate system are transformed into coordinates in the world coordinate system, and the coordinates of all pixels in the world coordinate system constitute a point cloud map; the point cloud map is filtered to obtain the orbit point cloud map.
4. The method for surveying construction of ballastless track for high-speed railway according to claim 3, characterized in that, The methods for obtaining the coordinate sequence of the inner points of the orbit and the orbital altitude sequence include: Based on the track point cloud map, slices are made along the track extension direction at preset slicing intervals. Point cloud data at the top surface of the track in each slice is extracted, and the arithmetic mean of the height coordinates is calculated based on the point cloud data at the top surface of the track. The obtained arithmetic mean is arranged in spatial order to obtain the track height sequence. Establish a reference coordinate system at the highest point of the top surface of the track in the track point cloud map. In the reference coordinate system, the X-axis is along the extension direction of the track, the Y-axis points to the inside of the track, and the Z-axis is vertically upward along the XY plane. A feature extraction plane is constructed by moving down the Z-axis by a preset distance. A sampling window is established along the extension direction of the track at a preset slice interval. The point cloud coordinates inside the track in the feature extraction plane are obtained through the sampling window, and the coordinate sequence of the points inside the track is obtained.