An orbit detection system and method based on adaptive matching and multi-source fusion

Abstract

This invention discloses a track detection system and method based on adaptive matching and multi-source fusion, belonging to the field of track detection technology. It solves the problems of weak anti-interference in existing profile matching, insufficient robustness of multi-source data fusion, and inaccurate defect tracing. The system includes a 3D camera acquiring profile and fastener images, laser inertial navigation and GNSS acquiring attitude positioning data, and a synchronous controller, processing computer, and odometer for synchronous triggering and data processing. The processing computer processes 3D camera data through adaptive weighted profile matching and fuses multi-source positioning data using federated Kalman filtering. Combined with track mapping, sleeper counting, and geometric parameter extraction modules, it realizes rail wear calculation, track parameter extraction, sleeper counting, and geometric parameter extraction. It accurately associates rail wear, track geometric parameters, and sleeper numbers, achieving precise positioning of track detection results and reducing detection mileage positioning errors.

Description

Technical Field

[0001] This invention relates to the field of track detection technology, and in particular to a track detection system and method based on adaptive matching and multi-source fusion. Background Technology

[0002] As the core infrastructure of railway transportation, the technical condition of railway tracks directly determines the safety, economy, and passenger comfort of train operation. With the continuous growth of railway network mileage and the continuous improvement of train operating speed and carrying capacity, the industry has placed higher demands on track inspection technology. There is an urgent need for inspection technology with higher detection accuracy, stronger anti-interference capabilities, and inspection results that can provide direct and effective guidance for track operation and maintenance.

[0003] While current non-contact track inspection technology has been applied, it still has many technical limitations and is difficult to adapt to the current inspection needs of railway tracks. Specifically, these problems manifest in three aspects: First, existing profile matching algorithms, which use fixed weight allocation and single template matching, are easily affected by factors such as rail wear, surface corrosion, and changes in ambient lighting, resulting in large profile matching errors. This makes it impossible to accurately calculate the relative pose between the inspection carrier and the rail, thus affecting the accuracy of rail wear calculations and track geometric parameter extraction. Second, in the process of multi-source sensor data fusion, the different inspection scenarios, such as open-air environments and tunnels, are not considered. The weights of data from each sensor are dynamically adjusted based on environmental characteristics. When a sensor fails in a specific scenario, it directly leads to insufficient accuracy in estimating the motion trajectory of the detection vehicle, creating potential errors for subsequent track mapping and parameter analysis. Sleeper technology often uses single-point laser or ultrasonic ranging combined with threshold judgment, resulting in low counting accuracy and susceptibility to environmental interference. Furthermore, a precise correlation between track defect data and sleeper numbers has not been established, making it difficult for maintenance personnel to quickly and accurately locate defects based on detection results. Consequently, targeted track fine-tuning and maintenance plans cannot be developed, reducing the efficiency and specificity of track maintenance.

[0004] Therefore, the industry urgently needs a non-contact track inspection technology that has high-precision profile matching, scene-adaptive multi-source data fusion, high-integration detection capabilities, and can accurately trace the source of defects, in order to solve the above-mentioned shortcomings of existing technologies and adapt to the development needs of high-quality operation and maintenance of railway tracks. Summary of the Invention

[0005] In view of this, this application provides an orbit detection system and method based on adaptive matching and multi-source fusion to address the shortcomings of the existing technology.

[0006] The first aspect of this application provides a track detection system based on adaptive matching and multi-source fusion, including a vehicle frame body, multiple sets of line structured light 3D cameras, a laser inertial navigation unit, a GNSS positioning module, a synchronization controller, a processing computer, and an odometer; The line structured light 3D camera is symmetrically arranged on both sides of the vehicle frame body, with the optical axis perpendicular to the axis of the rail to be measured, and is used to simultaneously collect the inner profile data of the rail and the image data of the fastener area. Both the laser inertial navigation unit and the GNSS positioning module are integrated in the middle of the vehicle frame. The laser inertial navigation unit is used to collect the carrier's motion attitude and speed data, and the GNSS positioning module is used to collect the carrier's real-time positioning data. The synchronization controller, the processing computer, and the odometer are all integrated at the bottom of the vehicle frame. The synchronization controller uses the coded mileage of the odometer as the trigger reference and outputs trigger pulses to synchronize and control the acquisition of multi-source data. The processing computer has built-in adaptive weighted profile matching module, federated Kalman filter fusion module, track mapping module, sleeper counting module and geometric parameter extraction module; The adaptive weighted profile matching module is used to perform anti-interference optimization, adaptive feature point extraction, and multi-template dynamic matching on the inner profile data of the rail, and outputs the optimal rotation and translation matrix between the standard profile and the measured profile. The Federal Kalman Filter Fusion Module is used to perform scene adaptive fusion of the output data of the laser inertial navigation unit, the output data of the odometer, and the output data of the GNSS positioning module to calculate the motion trajectory of the vehicle in the geodetic coordinate system. The track mapping module is used to calculate the true trajectory of the midpoint of the left and right rail tops based on the motion trajectory of the carrier in the geodetic coordinate system and the relative pose of the line structured light 3D camera and the rail. The sleeper counting module is used to count sleepers and establish a correspondence between defects and sleeper numbers based on the image data of the fastener area collected by the line structured light 3D camera. The geometric parameter extraction module is used to extract track geometric parameters based on the actual trajectory of the midpoints of the left and right rail tops.

[0007] In one possible implementation of the first aspect, the adaptive weighted profile matching module works as follows: Anti-interference optimization: Gaussian filtering was used to smooth the initially acquired inner profile data of the rail, and then morphological opening was used to remove interference points from the filtered inner profile data of the rail. Simultaneously, the ambient light sensor built into the line structured light 3D camera collects the light intensity, and based on the light intensity, the grayscale threshold is adaptively adjusted to correct the distortion of the inner profile data of the rail caused by the change in light intensity, thereby obtaining optimized inner profile data of the rail. Adaptive feature point extraction: Calculate the gradient change value of the optimized rail inner profile data, fit the corresponding rail inner profile curve based on the least squares method and obtain the curvature value by taking the derivative; Based on the abrupt changes in the gradient and curvature values, feature points in key areas of the rail head, rail web, and rail bottom are automatically identified, and interference points in worn and corroded areas are eliminated. Construct a weight calculation model and dynamically assign weights to feature points, using the following formula: For the first The weight values ​​for each feature point range from 0.1 to 0.9. For the first The gray-level variance value of each feature point For the first Gradient magnitude values ​​at each feature point; These are the weighting coefficients related to the gray-level variance of the feature points. These are the weighting coefficients for the gradient magnitude at the feature points; For the first The neighborhood consistency coefficient of each feature point ranges from 0 to 1. Dynamic matching of multiple templates: Three types of basic rail profile templates are preset: standard rail, lightly worn rail, and turnout area rail. Based on the feature points after dynamic weight allocation, the hausdorf distance between the inner profile data of the rail and the three types of basic rail profile templates is calculated and optimized, and the template with the highest similarity is automatically selected. A local iterative matching strategy is adopted, with a fixed iteration step size, to continuously optimize the rotation matrix and translation vector of the adaptation template, and finally output the optimal rotation and translation matrix of the standard contour and the measured contour.

[0008] In one possible implementation of the first aspect, the specific working process of the federated Kalman filter fusion module is as follows: Establish the system state equations and observation equations: The system state vector includes the position, velocity, attitude angle, and sensor zero bias of the carrier; the observation vector includes the motion attitude and velocity data output by the laser inertial navigation unit, the relative mileage data output by the odometer, and the real-time positioning data output by the GNSS positioning module. Dynamically assign data weights based on the scenario: The GNSS positioning module receives GNSS signal strength to identify whether the scene is an open-air scene or a tunnel scene. When it is an open-air scene, scene adaptive weights are assigned to the motion attitude and speed data, the relative mileage data and the real-time positioning data in the observation vector. When it is a tunnel scene, the real-time positioning data is masked and scene adaptive weights are assigned to the motion attitude and speed data and the relative mileage data. Federal Kalman filter fusion: The system uses a federated Kalman filter algorithm to fuse multi-source data, including motion attitude and velocity data, relative mileage data, and real-time positioning data, and iteratively updates the system state vector to calculate the vehicle's trajectory in the geodetic coordinate system.

[0009] In one possible implementation of the first aspect, the specific working process of the trajectory mapping module is as follows: The initial pose calibration of the line structured light 3D camera and the laser inertial navigation unit is completed based on the rail profile calibration plate, and a fixed transformation relationship between the camera coordinate system and the inertial navigation coordinate system is established. The fixed transformation relationship includes a rotation matrix and a translation vector. Combining the optimal rotation and translation matrix output by the adaptive weighted profile matching module, the real-time pose of the carrier relative to the left and right rails at each trajectory point is calculated. The real-time pose includes translation amount and rotation angle. Based on the fixed transformation relationship between the camera coordinate system and the inertial navigation coordinate system, the motion trajectory of the carrier in the geodetic coordinate system is transformed to the rail coordinate system, and then mapped to the real trajectory of the midpoint of the left and right rail tops, which serves as the real alignment of the rail.

[0010] In one possible implementation of the first aspect, the specific working process of the sleeper counting module is as follows: The image data of the fastener area is preprocessed, including grayscale conversion, histogram equalization, and noise suppression. A deep learning object detection model is used to identify fasteners in the preprocessed fastener area image, and the position coordinates and corresponding confidence scores of the fasteners are output. Based on the distribution characteristics and spacing consistency of the fasteners, a sliding window is used to calculate the average value of the sleeper spacing within a certain length. After removing duplicate detection results with consistent mileage and erroneous detection results with abnormal spacing, a continuous sleeper number is generated starting from the detection starting point. By combining the relative mileage data of the odometer after multi-source data fusion and calibration, the correspondence between track geometric parameters, rail wear data and corresponding sleeper numbers is established, and a detection data association table carrying sleeper numbers is generated and output.

[0011] In one possible implementation of the first aspect, the specific working process of the geometric parameter extraction module is as follows: Based on the actual trajectory of the midpoint of the left and right rail tops, and combined with the midpoint chord measurement method and parameter calculation model, the track geometric parameters are extracted. The track geometric parameters include track orientation, elevation, track gauge, level, superelevation, and triangular pit. Based on the actual trajectory of the midpoints of the left and right rail tops, the rail top coordinates P1 (X1,Y1,Z1), P2 (X2,Y2,Z2), and P3 (X3,Y3,Z3) are extracted; P2 is the midpoint of P1 and P3 along the track mileage direction. Based on P1, P2, and P3, the track direction and elevation parameters are calculated using the midpoint chord measurement method. The horizontal and superelevation parameters are calculated based on the difference in elevation between the left and right rail tops and the rail surface inclination. The formula is as follows: , The standard distance between the midpoints of the left and right rail tops. The roll attitude angle of the carrier; The parameters of the triangular pit are calculated based on the algebraic difference between the horizontal differences of the two cross-sections within the base length. The formula is as follows: , and These are the horizontal measurements of two cross sections within the base length; Tangents are drawn to the vertices of the actual trajectories of the midpoints of the left and right rails respectively. The tangents under the same cross section are shifted downward by a set value. The Euclidean distance between the intersection points of the left and right rail profile data is calculated to obtain the track gauge parameters.

[0012] In one possible implementation of the first aspect, the deep learning object detection model is the YOLOv8 model, and the confidence level for fastener recognition is ≥0.95; the size of the sliding window is adapted to the length of the sleeper.

[0013] In one possible implementation of the first aspect, the system adapts to operating environments including open-air mainline operation scenarios, tunnel operation scenarios, and turnout area operation scenarios.

[0014] In one possible implementation of the first aspect, the number of line structured light 3D cameras is 4; the GNSS positioning module provides RTK centimeter-level positioning; the odometer outputs 1 pulse for every 1 mm traveled, providing a mileage trigger reference for data synchronization and positioning.

[0015] A second aspect of this application provides a trajectory detection method based on adaptive matching and multi-source fusion, comprising: The frame body is mounted on the rail to be inspected, the attitude of the line structured light 3D camera is adjusted, the initial pose calibration of the line structured light 3D camera and the laser inertial navigation unit is completed through the rail profile calibration plate, the preset rail profile template and the pre-trained fastener detection deep learning target detection model are loaded, and the parameters are initialized. The synchronization controller receives the coded mileage output by the odometer in real time. When the coded mileage reaches or exceeds the mileage point in the predetermined mileage sequence generated at fixed spatial intervals for the first time, the synchronization controller outputs a trigger pulse, which synchronously triggers the structured light 3D camera to collect the inner profile data of the rail and the image data of the fastener area, the laser inertial navigation unit to collect the motion attitude and speed data of the carrier, and the GNSS positioning module to collect real-time positioning data, thereby realizing the spatiotemporal synchronization of multi-source detection data. The processing computer uses a federated Kalman filter fusion module to fuse laser inertial navigation data, odometry relative mileage data, and GNSS positioning data to calculate the trajectory of the vehicle in the geodetic coordinate system. An adaptive weighted profile matching algorithm is used to sequentially perform anti-interference optimization, feature point extraction, and multi-template dynamic matching on the inner profile data of the rail. The optimal rotation and translation matrix corresponding to the standard rail profile and the measured rail profile is output. Combined with the initial pose calibration results, the real-time pose of the carrier relative to the left and right rails is calculated. At the same time, based on the matching difference between the standard rail profile and the measured rail profile, the vertical wear and side wear of the rail are calculated as rail wear data. Based on the motion trajectory of the carrier in the geodetic coordinate system and the real-time position of the carrier relative to the left and right rails, the motion trajectory of the carrier is transformed into the real trajectory of the midpoint of the top of the left and right rails; using the midpoint chord measurement method and a preset parameter calculation model, the track orientation, elevation, gauge, level, superelevation and triangular pit geometric parameters of the track are extracted from the real trajectory. The image data of the fastener area acquired by the line structured light 3D camera is preprocessed, and the fastener targets in the image are identified by the deep learning target detection model to complete the automatic counting of sleepers and generate continuous sleeper numbers; combined with the relative mileage data of the odometer after multi-source data fusion and calibration, the correspondence between rail wear data, track geometric parameter anomaly data and corresponding sleeper numbers is established. By integrating data on rail wear, abnormal track geometry parameters, and sleeper numbers, a standardized track inspection report is generated, providing data support for track maintenance and fine-tuning operations.

[0016] Its beneficial effects are as follows: This invention discloses a non-contact intelligent track inspection system, comprising a vehicle frame, multiple sets of line structured light 3D cameras, a laser inertial navigation unit, a GNSS positioning module, a synchronization controller, a processing computer, and an odometer. The 3D cameras are symmetrically arranged on both sides of the vehicle frame, simultaneously acquiring images of the inner profile of the rails and the fastener area; the laser inertial navigation unit and GNSS positioning module are integrated in the middle of the vehicle frame, acquiring the carrier's attitude, velocity, and real-time positioning data; the synchronization controller, processing computer, and odometer are located at the bottom of the vehicle frame, using the odometer's encoded mileage as a trigger reference to achieve synchronous acquisition of multi-source data. The processing computer has five built-in algorithm modules, which process 3D camera data through adaptive weighted profile matching, fuse multi-source positioning data through federated Kalman filtering, and combine with track mapping, sleeper counting, and geometric parameter extraction modules to complete rail wear calculation, track geometric parameter extraction, sleeper counting, and accurate correlation between defects and sleeper numbers. This invention enables non-contact, high-precision track inspection, effectively improving the anti-interference capability of profile matching and the robustness of multi-source data fusion. It provides accurate fault location and tracing, and can be adapted to inspection needs in various scenarios such as open-air, tunnel, and turnout. The system has high integration and strong inspection stability, providing accurate data support for track operation and maintenance fine-tuning, and significantly improving track inspection efficiency and operation and maintenance guidance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the composition of an orbit detection system based on adaptive matching and multi-source fusion provided in an embodiment of this application; Figure 2 This is a schematic diagram of a trajectory detection method based on adaptive matching and multi-source fusion provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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. Without further limitation, 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 the element.

[0021] Example 1 Existing non-contact track inspection technologies have the following limitations: First, the profile matching algorithm uses fixed weight allocation and single template matching, which is easily affected by rail wear, corrosion, and changes in lighting, resulting in large matching errors and making it impossible to accurately calculate the relative pose of the carrier and the rail. Second, the data weights are not dynamically adjusted for different scenarios during multi-source data fusion, resulting in insufficient accuracy in carrier trajectory estimation. Third, sleeper counting often uses single-point laser or ultrasonic ranging combined with threshold counting, which has low accuracy and is easily affected by environmental interference. At the same time, no accurate correlation is established between defects and sleeper numbers, making it difficult for maintenance personnel to quickly locate defect locations and formulate targeted fine-tuning plans.

[0022] Therefore, this application provides an orbit detection system based on adaptive and multi-source fusion, such as... Figure 1 As shown, it includes the vehicle frame body, multiple sets of line structured light 3D cameras, laser inertial navigation unit, GNSS positioning module, synchronization controller, processing computer and odometer; The line structured light 3D camera is symmetrically arranged on both sides of the vehicle frame body, with the optical axis perpendicular to the axis of the rail to be measured, and is used to simultaneously collect the inner profile data of the rail and the image data of the fastener area. Both the laser inertial navigation unit and the GNSS positioning module are integrated in the middle of the vehicle frame. The laser inertial navigation unit is used to collect the carrier's motion attitude and speed data, and the GNSS positioning module is used to collect the carrier's real-time positioning data. The synchronization controller, the processing computer, and the odometer are all integrated at the bottom of the vehicle frame. The synchronization controller uses the coded mileage of the odometer as the trigger reference and outputs trigger pulses to synchronize and control the acquisition of multi-source data. The processing computer has built-in adaptive weighted profile matching module, federated Kalman filter fusion module, track mapping module, sleeper counting module and geometric parameter extraction module; The adaptive weighted profile matching module is used to perform anti-interference optimization, adaptive feature point extraction, and multi-template dynamic matching on the inner profile data of the rail, and outputs the optimal rotation and translation matrix between the standard profile and the measured profile. The Federal Kalman Filter Fusion Module is used to perform scene adaptive fusion of the output data of the laser inertial navigation unit, the output data of the odometer, and the output data of the GNSS positioning module to calculate the motion trajectory of the vehicle in the geodetic coordinate system. The track mapping module is used to calculate the true trajectory of the midpoint of the left and right rail tops based on the motion trajectory of the carrier in the geodetic coordinate system and the relative pose of the line structured light 3D camera and the rail. The sleeper counting module is used to count sleepers and establish a correspondence between defects and sleeper numbers based on the image data of the fastener area collected by the line structured light 3D camera. The geometric parameter extraction module is used to extract track geometric parameters based on the actual trajectory of the midpoints of the left and right rail tops.

[0023] The track detection system in this embodiment consists of a hardware module and an algorithm module. The hardware module is responsible for the synchronous acquisition and transmission of multi-source data, while the algorithm module is responsible for the processing, fusion and analysis of the data. Together, they achieve fully automated detection from raw data acquisition to output of detection results.

[0024] The system hardware modules include the vehicle frame, multiple sets of line structured light 3D cameras, a laser inertial navigation unit, a GNSS positioning module, a synchronization controller, a processing computer, and an odometer. The algorithm module, integrated within the processing computer, includes an adaptive weighted profile matching module, a federated Kalman filter fusion module, a track mapping module, a sleeper counting module, and a geometric parameter extraction module. The layout and connection relationships of each module are as follows: the line structured light 3D cameras are symmetrically arranged on both sides of the vehicle frame; the laser inertial navigation unit and GNSS positioning module are integrated in the middle of the vehicle frame; and the synchronization controller, processing computer, and odometer are integrated at the bottom of the vehicle frame. The synchronization controller connects to each hardware component and the processing computer via cables to achieve synchronous triggering and transmission of multi-source data. The processing computer centrally processes the collected raw data through its built-in algorithm module and outputs the detection results.

[0025] Hardware module working principle: Frame body: The system hardware mounting carrier and motion platform provide stable physical support for all detection components, ensuring that the relative positions of each component are fixed during the detection process, which is the basis for consistent data acquisition.

[0026] Line structured light 3D camera: The system's core visual data acquisition component has dual functions: acquiring three-dimensional data of the inner profile of the rail and acquiring two-dimensional images of the fastener area. It is the source of raw data for rail wear calculation and sleeper counting.

[0027] In this embodiment, there are 4 cameras, symmetrically arranged on both sides of the frame (2 cameras on each side). The optical axis is strictly perpendicular to the rail axis to ensure that the collected profile data has no angular distortion with the actual rail profile. The cameras on both sides correspond to the left and right rails respectively, realizing independent and synchronous acquisition of left and right rail data. The acquisition frequency is consistent with the trigger frequency of the synchronization controller. Data acquisition is only performed when the synchronization controller outputs a trigger pulse to ensure the spatiotemporal synchronization of the acquired profile data, fastener image data and mileage, inertial navigation and GNSS data.

[0028] Laser inertial navigation unit: The system's core component for detecting the carrier's motion attitude and state provides high-frequency, high-reliability attitude and velocity data for carrier trajectory estimation and pose calculation. It integrates a 3-axis laser gyroscope and a 3-axis quartz accelerometer, totaling six detection axes, enabling full-dimensional detection of the carrier's spatial motion.

[0029] Mounted at the geometric center of the vehicle frame, it is aligned with the vehicle's center of mass as closely as possible to minimize attitude and velocity detection errors caused by eccentricity. The acquisition action is triggered by a synchronous controller, which outputs attitude angle and velocity data at high frequency and transmits them to the processing computer in real time, providing core data for federated Kalman filter fusion. Even in scenarios where GNSS signals are unavailable (such as tunnels), it can continuously detect the vehicle's motion state on its own, making it a core reference component for trajectory extrapolation.

[0030] GNSS positioning module: The system's outdoor spatial positioning component provides an absolute position reference for calculating the trajectory of a vehicle in open-air scenarios, compensating for the "integral drift" defect of the laser inertial navigation unit and improving the long-term accuracy of trajectory calculation. It supports BeiDou + GPS dual-mode positioning and adopts RTK (real-time dynamic) positioning technology; at the same time, the module can detect GNSS signal strength in real time and transmit the signal strength data to the processing computer, providing a basis for scene recognition.

[0031] The laser inertial navigation unit is integrated into the same area of ​​the vehicle frame to ensure that the two are in close spatial position, reducing the fusion error of positioning and attitude data caused by position deviation. In open-air scenarios, high-frequency positioning data is output and fused with laser inertial navigation and odometer data. In tunnel scenarios, because satellite signals are blocked and the signal strength is below the threshold, the processing computer automatically blocks the output data and no longer participates in the fusion.

[0032] Odometer: The system's relative mileage detection and synchronous triggering reference components provide a spatial triggering reference for synchronous data acquisition and relative mileage data for carrier trajectory estimation. A photoelectric odometer is used, mechanically connected to the vehicle's wheels. As the wheels rotate, the odometer's encoder disk rotates synchronously. In this embodiment, the odometer outputs one pulse for every 1mm traveled, achieving millimeter-level relative mileage detection.

[0033] Synchronous controller: The multi-source data synchronization triggering core component of the system is the key to ensuring the spatiotemporal consistency of data collected by each hardware module, realizing unified triggering and synchronous acquisition; the triggering response time is short, ensuring the synchronization of acquisition actions of each component.

[0034] Handling computers: The system's data processing, algorithm calculation, and core control center are responsible for receiving, processing, fusing, and outputting results from multiple data sources. Equipped with a high-performance industrial processor and a large-capacity storage module, it establishes high-speed data transmission with the synchronization controller and various acquisition components via cables. It receives all data triggered by the synchronization controller in real time (profile and image data from the 3D camera, attitude and velocity data from the laser inertial navigation system, positioning data from GNSS, and relative mileage data from the odometer), and stores the raw data locally. It has five built-in core algorithm modules that process and calculate the raw data sequentially according to preset algorithm flows. From multi-source data fusion to infer the carrier trajectory, to profile matching to calculate relative pose, to track mapping to extract geometric parameters, to sleeper counting to establish defect correlations, and finally, it integrates all processing results to generate a standardized track inspection report.

[0035] It possesses powerful parallel computing capabilities and data processing speed, enabling real-time processing of high-frequency, massive amounts of multi-source detection data; it supports flexible upgrades and parameter adjustments of algorithm modules to adapt to the detection needs of different track scenarios; and it has high reliability, allowing it to operate stably in the complex environment of railway sites.

[0036] How the algorithm module works: Adaptive weighted profile matching module: The core module for rail profile data processing solves the problems of weak anti-interference ability and low matching accuracy in traditional profile matching. It achieves high-precision matching between the measured rail profile and the standard profile, and outputs the optimal rotation and translation matrix between the two, providing the core data foundation for the relative pose calculation of the carrier and the rail and the calculation of rail wear.

[0037] A three-level progressive processing flow of "anti-interference optimization, adaptive feature point extraction, and multi-template dynamic matching" is adopted to refine the 3D point cloud data of the inner profile of the rail acquired by the 3D camera: For anti-interference optimization, the original profile data is first smoothed by Gaussian filtering to suppress high-frequency interference caused by rail surface noise and measurement jitter. Then, morphological opening operation (corrosion followed by expansion) is used to remove impurities and sharp interference points formed by rust on the rail surface, restoring the true shape of the profile. At the same time, based on the light intensity data of the ambient light sensor built into the 3D camera, the grayscale threshold is adaptively adjusted to correct the profile data distortion caused by excessively strong or weak light, resulting in optimized inner profile data of the rail.

[0038] Adaptive feature point extraction first calculates the gradient change value of the optimized profile data, then fits the profile curve using the least squares method and obtains the curvature value through differentiation. Based on the abrupt changes in gradient and curvature, stable feature points in key areas of the rail head, rail web, and rail bottom are automatically identified, while invalid interference points in worn and corroded areas are removed. Secondly, a dynamic weight calculation model is constructed to assign weights to each feature point (overall weight range 0.1~0.9), using the following formula: in, 0.3 It is 0.4. The gray-level variance of feature points (reflecting stability). This represents the gradient magnitude (reflecting significance). The neighborhood consistency coefficient is 0-1, reflecting reliability. Furthermore, the weights of stable feature points at the rail base are 0.7-0.9, those at the rail web are 0.4-0.6, and those at the rail head (easily worn areas) are 0.1-0.3, thus tilting the matching process towards feature points with high reliability.

[0039] Multi-template dynamic matching is implemented, with three basic profile templates preset: standard rail, lightly worn rail, and turnout area rail. Based on feature points with dynamically assigned weights, the Hausdorff distance between the measured profile and each template is calculated (quantifying the similarity between the two profiles), and the template with the highest similarity is automatically selected. A local iterative matching strategy is adopted, with a minimum iteration step size of 0.01 mm, to continuously optimize the rotation matrix and translation vector of the template, maximizing the overlap between the measured profile and the template, and finally outputting the optimal rotation and translation matrix between the standard profile and the measured profile.

[0040] Abandoning the traditional fixed weight and single template matching method, it realizes dynamic allocation of feature point weights and adaptive template selection. Combined with multi-dimensional anti-interference processing, it significantly improves the anti-interference ability and accuracy of profile matching.

[0041] Federal Kalman filter fusion module: The core module of multi-source positioning data fusion solves the problems of poor adaptability of traditional data fusion scenarios and low trajectory estimation accuracy after sensor failure. It realizes scene adaptive fusion of multi-source data such as laser inertial navigation, odometer, and GNSS, and calculates the high-precision motion trajectory of the carrier in the geodetic coordinate system, providing a position reference for subsequent trajectory mapping.

[0042] The computational process of "system model establishment, dynamic scene weight allocation, and federated Kalman filter iteration" is adopted to fuse the attitude / velocity data of laser inertial navigation, the relative mileage data of odometry, and the positioning data of GNSS. The system model is established, and the system state equation and observation equation are constructed. The system state vector includes the position, velocity, attitude angle and sensor zero bias of the carrier; the observation vector includes the motion attitude and velocity data output by the laser inertial navigation unit, the relative mileage data output by the odometer, and the real-time positioning data output by the GNSS positioning module.

[0043] Scene dynamic weight allocation automatically identifies the detection scene (open-air / tunnel) based on the GNSS signal strength collected by the GNSS positioning module, and assigns differentiated weights to different sensor data: In open-air scenes, the GNSS signal is stable, and GNSS data is assigned a weight of 0.4-0.5, laser inertial navigation data 0.3-0.4, and odometer data 0.1-0.2; In tunnel scenes, the GNSS signal is blocked, and GNSS data is directly blocked, and laser inertial navigation data is assigned a weight of 0.6-0.7, and odometer data 0.3-0.4.

[0044] The federated Kalman filter iterative method adopts a federated filtering architecture, which divides multi-source data into a laser inertial navigation-odometry sub-filter and a GNSS sub-filter (disabled in tunnel scenarios). Each sub-filter independently performs state prediction and observation update based on the assigned weights. The main filter collects the calculation results of each sub-filter, performs global fusion and state correction, iteratively updates the system state vector, and finally calculates the high-precision, continuous motion trajectory of the vehicle in the geodetic coordinate system.

[0045] By introducing a scene-adaptive weight allocation mechanism and combining it with the federated filtering architecture of "independent operation of sub-filters + global fusion of main filters", the problem of single sensor failure (such as no GNSS signal in tunnels) is effectively solved, and the scene adaptability and accuracy of trajectory estimation are improved.

[0046] Track mapping module: The module for converting the carrier trajectory to the actual rail alignment enables precise conversion from the carrier's geodetic coordinate system trajectory to the actual trajectory of the midpoints of the left and right rail tops, providing a direct data basis for extracting track geometric parameters.

[0047] The computational process of "initial pose calibration, real-time pose calculation, and trajectory mapping transformation" is adopted. Combining hardware calibration data and the calculation results of previous modules, the multi-coordinate system transformation and mapping of the trajectory are completed. Initial pose calibration: Before testing, the initial pose calibration of the line structured light 3D camera and the laser inertial navigation unit is completed using a rail profile calibration plate. A fixed transformation relationship between the camera coordinate system and the inertial navigation coordinate system is established. This relationship remains unchanged during the testing process, providing a basic reference for subsequent coordinate transformations.

[0048] Real-time pose calculation, combined with the optimal rotation and translation matrix output by the adaptive weighted profile matching module (the pose of the measured profile relative to the standard profile), and the above fixed transformation relationship, uses a coordinate transformation algorithm to calculate the real-time pose of the carrier relative to the left and right rails at each trajectory point, including translation and rotation angles, reflecting the changes in the position and attitude of the carrier relative to the rails during the detection process.

[0049] The trajectory mapping transformation first utilizes the fixed transformation relationship between the camera coordinate system and the inertial navigation coordinate system to transform the carrier's geodetic coordinate system motion trajectory output by the federated Kalman filter fusion module to the camera coordinate system. Then, combined with the real-time pose of the carrier relative to the rail, the carrier trajectory in the camera coordinate system is further transformed to the rail coordinate system, and finally mapped to the real trajectory of the midpoint of the left and right rail tops (i.e., the real line shape of the rail). This trajectory is the core data source for the extraction of track geometric parameters.

[0050] By employing the dual constraints of "initial calibration and real-time pose calculation," precise conversion between multiple coordinate systems is achieved, transforming the absolute motion trajectory of the carrier into the actual alignment of the rail, thus eliminating the influence of carrier motion deviation on track parameter extraction.

[0051] Sleeper counting module: The core module for accurately linking sleeper counting with defects solves the problems of low accuracy and difficulty in tracing defects in traditional sleeper counting. Based on fastener area images acquired by a 3D camera, it achieves high-precision automatic counting of sleepers and establishes a correspondence between track defect data and sleeper numbers, providing accurate defect location information for track operation and maintenance.

[0052] The computational process of "image preprocessing, deep learning fastener recognition, automatic sleeper counting, and defect association mapping" is used to process the two-dimensional image data of the fastener area acquired by the 3D camera. Image preprocessing involves converting the original fastener area image to grayscale, transforming the color image into a grayscale image; histogram equalization is used to enhance the contrast between the fastener and the background; and median filtering is employed to suppress image noise, eliminating interference for subsequent fastener recognition and improving recognition accuracy.

[0053] Deep learning is used for fastener recognition. A pre-trained YOLOv8 target detection model (trained on a large dataset of fastener images from various scenarios) is loaded to detect fastener targets on the pre-processed images. The model outputs the position coordinates, bounding box, and confidence score of each fastener. A confidence score threshold of ≥0.95 is set to eliminate misidentified interference targets and ensure the accuracy of fastener recognition.

[0054] Automatic sleeper counting is based on the distribution characteristics of rail fasteners (each sleeper corresponds to a fixed number of fasteners) and the consistency of spacing (the standard fastener spacing corresponds to the sleeper spacing). It adopts a sliding window counting method (the window size is adapted to the sleeper length) to calculate the average sleeper spacing within a 20-meter length, remove duplicate detection results with the same mileage and erroneous detection results with abnormal spacing, and generate a continuous sleeper number that increases from the detection starting point, effectively controlling counting deviation.

[0055] The defect association mapping, combined with the relative mileage information of the odometer after multi-source data fusion and calibration, determines the mileage segment where defects such as rail wear and abnormal track geometric parameters are located. Then, the mileage segment is accurately associated with the corresponding sleeper number, and the detection data association table of "defect type-defect degree-corresponding sleeper number-mileage location" is output to achieve accurate source tracing of defects.

[0056] By combining YOLOv8 deep learning object detection with sliding window precision counting, the traditional laser / ultrasonic ranging counting method is replaced, improving the accuracy of sleeper counting. At the same time, a direct correlation is established between defects and sleeper numbers, enabling sleeper-level precise location of defects and significantly improving the targeted nature of track maintenance.

[0057] Geometric parameter extraction module: The core geometric parameters extraction module extracts the core geometric parameters of the track based on the actual trajectories of the midpoints of the left and right track tops output by the track mapping module, combined with a professional track detection algorithm.

[0058] Using the midpoint chord measurement method as the core, combined with a pre-set parameter calculation model, five major track geometric parameters—track orientation, elevation, track gauge, horizontal and superelevation, and triangular crater—are extracted from the actual trajectories of the midpoints of the left and right track tops. Track alignment and elevation are calculated by extracting three consecutive track top coordinates P1(X1,Y1,Z1), P2(X2,Y2,Z2), and P3(X3,Y3,Z3) from the actual trajectory of the midpoints of the left and right track tops. P2 is the midpoint of P1 and P3 along the track mileage direction. Based on P1, P2, and P3, the track alignment and elevation parameters are calculated using the midpoint chord method. Track alignment is the deviation of the track alignment on the horizontal plane, calculated from the deviation of the plane coordinates (X,Y) of P1, P2, and P3. Elevation is the deviation of the track alignment on the vertical plane, calculated from the deviation of the elevation coordinates (Z) of P1, P2, and P3.

[0059] The horizontal and superelevation parameters are calculated based on the difference in elevation between the left and right rail tops and the rail surface inclination. The formula is as follows: , The standard distance between the midpoints of the left and right rail tops. The horizontal axis represents the roll attitude angle of the carrier; the horizontal axis represents the left and right rail top elevation deviations on a straight track; and the superelevation represents the rail top elevation difference on a curved track set to counteract centrifugal force. Both are calculated using this formula.

[0060] The parameters of the triangular pit are calculated based on the algebraic difference between the horizontal differences of two cross-sections within the base length. The formula is as follows: , and The horizontal measurement values ​​of two sections within the base length; the triangular pit reflects the degree of planar distortion of the track within the same base length and is an important parameter for evaluating track stability.

[0061] The track gauge is calculated by drawing tangents to the vertices of the actual trajectories of the midpoints of the left and right rails, shifting the tangents at the same cross-section downwards by a set value (16mm, the standard position for rail inspection), and calculating the Euclidean distance between the intersections of the left and right rail profile data. This distance is the actual track gauge value, which accurately reflects the lateral spacing deviation of the rails.

[0062] Based on the actual rail alignment parameters, and combined with the midpoint chord measurement method and quantitative calculation model of the rail inspection industry standard, the automated and high-precision extraction of rail geometric parameters is achieved, and the parameter results are consistent with the actual on-site inspection needs.

[0063] The system in this embodiment does not work in isolation, but forms a closed-loop collaborative system of hardware acquisition, synchronous triggering, algorithm processing, and result output. The core collaborative logic is as follows: The odometer detects relative mileage in real time, providing a spatial reference for synchronous triggering; The synchronization controller outputs synchronization trigger pulses based on the odometer mileage to ensure that the 3D camera, laser inertial navigation and GNSS collect data at the same time point, so as to realize the spatiotemporal synchronization of multi-source data; 3D camera, laser inertial navigation, GNSS, and odometry collect visual, attitude, positioning, and odometer data respectively, which are transmitted to the processing computer via cable to provide multi-dimensional raw detection data for the algorithm module; The computer's five built-in algorithm modules work in sequence: the federated Kalman filter fusion module processes multi-source positioning data and outputs a high-precision trajectory in the carrier's geodetic coordinate system; the adaptive weighted profile matching module processes 3D camera profile data and outputs the optimal rotation and translation matrix; the track mapping module combines the above trajectory and matrix to calculate the true trajectory of the left and right rail top midpoints; the geometric parameter extraction module extracts track geometric parameters from the true trajectory, and the adaptive weighted profile matching module simultaneously calculates rail wear data; the sleeper counting module processes 3D camera fastener images, completes sleeper counting, and establishes a correlation between defects and sleeper numbers. The computer integrates rail wear data, track geometry parameters, and defect-sleeper correlation data to generate a standardized inspection report containing data tables, trajectory curves, defect distribution maps, and fine-tuning suggestions, providing accurate data support for track operation and maintenance and fine-tuning.

[0064] This embodiment provides an orbit detection system based on adaptive matching and multi-source fusion, which makes at least the following technical contributions compared to existing technologies: The detection accuracy has been significantly improved. By using dynamic weighting of feature points and multi-template matching, the problem of weak anti-interference in traditional profile matching has been solved, and the accuracy of rail wear calculation and pose calculation has been greatly optimized. The multi-source data scene adaptive fusion strategy ensures high accuracy of carrier trajectory estimation in different environments such as open air and tunnels, laying the foundation for track parameter extraction.

[0065] It has strong scene adaptability, compact system hardware layout, and algorithm module supports adaptive adjustment of parameters in multiple scenes. It can seamlessly adapt to complex track detection needs such as open trunk lines, tunnels, and turnouts without the need for additional hardware structure adjustments.

[0066] It provides excellent operation and maintenance guidance, achieving high-precision counting of sleepers through deep learning target detection. It establishes a precise correlation between defects such as rail wear and abnormal geometric parameters and sleeper numbers, allowing operation and maintenance personnel to quickly locate the defects and formulate targeted fine-tuning plans, thus significantly improving operation and maintenance efficiency.

[0067] With excellent integration and stability, the 3D camera combines contouring and image acquisition functions, simplifying the system structure and reducing equipment costs. Multi-dimensional anti-interference processing (filtering and noise reduction, adaptive lighting adjustment, etc.) and hardware co-design ensure stable and reliable detection process and strong data output consistency.

[0068] Example 2 Based on the orbit detection system based on adaptive matching and multi-source fusion provided in Embodiment 1 of this application, correspondingly, Embodiment 2 of this application also provides an orbit detection method based on adaptive matching and multi-source fusion, such as... Figure 2 As shown, it includes: The frame body is mounted on the rail to be inspected, the attitude of the line structured light 3D camera is adjusted, the initial pose calibration of the line structured light 3D camera and the laser inertial navigation unit is completed through the rail profile calibration plate, the preset rail profile template and the pre-trained fastener detection deep learning target detection model are loaded, and the parameters are initialized. The synchronization controller receives the coded mileage output by the odometer in real time. When the coded mileage reaches or exceeds the mileage point in the predetermined mileage sequence generated at fixed spatial intervals for the first time, the synchronization controller outputs a trigger pulse, which synchronously triggers the structured light 3D camera to collect the inner profile data of the rail and the image data of the fastener area, the laser inertial navigation unit to collect the motion attitude and speed data of the carrier, and the GNSS positioning module to collect real-time positioning data, thereby realizing the spatiotemporal synchronization of multi-source detection data. The processing computer uses a federated Kalman filter fusion module to fuse laser inertial navigation data, odometry relative mileage data, and GNSS positioning data to calculate the trajectory of the vehicle in the geodetic coordinate system. An adaptive weighted profile matching algorithm is used to sequentially perform anti-interference optimization, feature point extraction, and multi-template dynamic matching on the inner profile data of the rail. The optimal rotation and translation matrix corresponding to the standard rail profile and the measured rail profile is output. Combined with the initial pose calibration results, the real-time pose of the carrier relative to the left and right rails is calculated. At the same time, based on the matching difference between the standard rail profile and the measured rail profile, the vertical wear and side wear of the rail are calculated as rail wear data. Based on the motion trajectory of the carrier in the geodetic coordinate system and the real-time position of the carrier relative to the left and right rails, the motion trajectory of the carrier is transformed into the real trajectory of the midpoint of the top of the left and right rails; using the midpoint chord measurement method and a preset parameter calculation model, the track orientation, elevation, gauge, level, superelevation and triangular pit geometric parameters of the track are extracted from the real trajectory. The image data of the fastener area acquired by the line structured light 3D camera is preprocessed, and the fastener targets in the image are identified by the deep learning target detection model to complete the automatic counting of sleepers and generate continuous sleeper numbers; combined with the relative mileage data of the odometer after multi-source data fusion and calibration, the correspondence between rail wear data, track geometric parameter anomaly data and corresponding sleeper numbers is established. By integrating data on rail wear, abnormal track geometry parameters, and sleeper numbers, a standardized track inspection report is generated, providing data support for track maintenance and fine-tuning operations.

[0069] Example 3 This embodiment provides a track detection system and method based on adaptive matching and multi-source fusion, applied to the detection scenario of open-air mainline tracks, as detailed below: System configuration: Line structured light 3D camera with a line frequency of 7kHz, X-axis accuracy of 0.1mm, and Z-axis accuracy of 0.01mm; Laser inertial navigation unit (IMU) adopts high-precision fiber optic inertial navigation with an output frequency of 200Hz; GNSS positioning module supports BeiDou + GPS dual-mode positioning; Odometer outputs 1 pulse for every 1mm traveled; Processing computer is equipped with a high-performance processor and has a built-in YOLOv8 pre-trained model.

[0070] Testing process: The frame body is placed on the mainline rail to be tested (standard rail = 1435mm). The initial pose calibration of the 3D camera and laser inertial navigation unit is completed by using the rail profile calibration plate. The standard rail template and YOLOv8 fastener detection model are loaded, and the algorithm parameters are initialized. The synchronous controller triggers each component to collect data at 1mm mileage intervals. The 3D camera synchronously collects the inner profile data of the rail and the image data of the fastener area. The laser inertial navigation unit collects the carrier's attitude and speed data. The GNSS positioning module collects real-time positioning data. The odometer collects relative mileage data. The Federal Kalman filter fusion module allocates weights according to the open-air scene (GNSS 0.45, laser inertial navigation 0.35, odometry 0.2) to fuse data and calculate the vehicle's motion trajectory; An adaptive weighted profile matching algorithm is used to extract feature points of the rail head, rail web, and rail base, and dynamically assign weights (0.8 for rail base, 0.5 for rail web, and 0.2 for rail head). A standard rail template is selected for local iterative matching (step size 0.01 mm). After anti-interference processing, the rotation and translation matrix is ​​output, and the vertical and side wear data of the rail are calculated. The track mapping module combines initial calibration data with real-time pose to map the carrier trajectory to the actual trajectory of the midpoint of the left and right rail tops; the midpoint chord measurement method is used to extract parameters such as track orientation, elevation, track gauge, level and superelevation, and triangular pit. The parameter results and standard values ​​deviate from the requirements of high-precision detection. After preprocessing, the fastener area images captured by the 3D camera are used to identify the fasteners using the YOLOv8 model (confidence level ≥ 0.96). The sleeper numbers 1-500 are obtained using the sliding window counting method. The wear and tear defects are associated with sleeper numbers 120-125 by combining the mileage data. The output includes a test report containing wear data, geometric parameters, and a defect-sleeper correlation table.

[0071] Example 4 This embodiment provides a track detection system and method based on adaptive matching and multi-source fusion, applied to a track detection scenario inside a tunnel, as detailed below: The system configuration is shown in Example 3.

[0072] Testing process: The deployment and data acquisition steps are the same as in Example 3. The GNSS signal strength is weak in the tunnel. The system automatically identifies the tunnel scene and turns off the GNSS sub-filter. The federal Kalman filter fusion module allocates fused data according to the tunnel scene weights (0.65 for laser inertial navigation and 0.35 for odometer) to calculate the vehicle's motion trajectory; The adaptive weighted profile matching algorithm addresses the problem of insufficient lighting in tunnels by adaptively adjusting the grayscale threshold to 80 and outputting a rotation and translation matrix after anti-interference processing to calculate the vertical and lateral wear data of the rail. The track mapping module maps the carrier trajectory to the real trajectory of the midpoint of the left and right rail tops, and extracts parameters such as track orientation, elevation, track gauge, horizontal and superelevation, and triangular pit. The parameter extraction results are stable and reliable. A 3D camera captures images of the fastener area under tunnel lighting. After histogram equalization to enhance contrast, the YOLOv8 model identifies the fasteners (confidence ≥ 0.95), completes sleeper counting (numbers 300-800), and associates wear defects with sleeper numbers 300-308. Output track inspection reports inside the tunnel, accurately matching the location of defects to the sleeper number.

[0073] Example 5 This embodiment provides a track detection system and method based on adaptive matching and multi-source fusion, applied to track detection scenarios in turnout areas, as detailed below: The system configuration is shown in Example 3.

[0074] Testing process: During initialization, the system automatically identifies the turnout area, loads the turnout rail template, and completes the initial position calibration. Simultaneous data acquisition: The rail profile in the turnout area is complex, and the 3D camera accurately acquires the inner profile data and the fastener images in the turnout area. The Federal Kalman Filter Fusion Module allocates fused data according to the weights of the open-air scene to calculate the vehicle's motion trajectory; An adaptive weighted profile matching algorithm is used to extract special feature points of the turnout rail (frog region), dynamically assign weights (rail base 0.9, rail web 0.6, rail head 0.3), output rotation and translation matrix after local iterative matching, and calculate the vertical wear data of the frog region. The track mapping module maps the carrier trajectory to the real trajectory of the midpoint of the left and right rail tops in the turnout area, and extracts the track gauge and superelevation parameters. A 3D camera captures images of fasteners in the turnout area, and a YOLOv8 model identifies special fasteners (confidence ≥ 0.95), completes sleeper counting (numbers 600-700), and associates turnout wear with sleeper numbers 600-603; Output turnout area inspection reports to provide data support for turnout fine-tuning.

[0075] Examples 3 to 5 demonstrate that the track detection system and method based on adaptive matching and multi-source fusion provided by the present invention can achieve stable and reliable track detection in various scenarios such as open-air environments, tunnels, and turnouts, with accurate defect location, and can effectively meet the needs of track operation and maintenance and fine-tuning.

[0076] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computing software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0077] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0078] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A track detection system based on adaptive matching and multi-source fusion, characterized in that, It includes the vehicle frame, multiple sets of line structured light 3D cameras, laser inertial navigation unit, GNSS positioning module, synchronization controller, processing computer and odometer; The line structured light 3D camera is symmetrically arranged on both sides of the vehicle frame body, with the optical axis perpendicular to the axis of the rail to be measured, and is used to simultaneously collect the inner profile data of the rail and the image data of the fastener area. Both the laser inertial navigation unit and the GNSS positioning module are integrated in the middle of the vehicle frame. The laser inertial navigation unit is used to collect the carrier's motion attitude and speed data, and the GNSS positioning module is used to collect the carrier's real-time positioning data. The synchronization controller, the processing computer, and the odometer are all integrated at the bottom of the vehicle frame. The synchronization controller uses the coded mileage of the odometer as the trigger reference and outputs trigger pulses to synchronize and control the acquisition of multi-source data. The processing computer has built-in adaptive weighted profile matching module, federated Kalman filter fusion module, track mapping module, sleeper counting module and geometric parameter extraction module; The adaptive weighted profile matching module is used to perform anti-interference optimization, adaptive feature point extraction, and multi-template dynamic matching on the inner profile data of the rail, and outputs the optimal rotation and translation matrix between the standard profile and the measured profile. The Federal Kalman Filter Fusion Module is used to perform scene adaptive fusion of the output data of the laser inertial navigation unit, the output data of the odometer, and the output data of the GNSS positioning module to calculate the motion trajectory of the vehicle in the geodetic coordinate system. The track mapping module is used to calculate the true trajectory of the midpoint of the left and right rail tops based on the motion trajectory of the carrier in the geodetic coordinate system and the relative pose of the line structured light 3D camera and the rail. The sleeper counting module is used to count sleepers and establish a correspondence between defects and sleeper numbers based on the image data of the fastener area collected by the line structured light 3D camera. The geometric parameter extraction module is used to extract track geometric parameters based on the actual trajectory of the midpoints of the left and right rail tops.

2. The orbit detection system based on adaptive matching and multi-source fusion according to claim 1, characterized in that, The specific working process of the adaptive weighted profile matching module is as follows: Anti-interference optimization: Gaussian filtering was used to smooth the initially acquired inner profile data of the rail, and then morphological opening was used to remove interference points from the filtered inner profile data of the rail. Simultaneously, the ambient light sensor built into the line structured light 3D camera collects the light intensity, and based on the light intensity, the grayscale threshold is adaptively adjusted to correct the distortion of the inner profile data of the rail caused by the change in light intensity, thereby obtaining optimized inner profile data of the rail. Adaptive feature point extraction: Calculate the gradient change value of the optimized rail inner profile data, fit the corresponding rail inner profile curve based on the least squares method and obtain the curvature value by taking the derivative; Based on the abrupt changes in the gradient and curvature values, feature points in key areas of the rail head, rail web, and rail bottom are automatically identified, and interference points in worn and corroded areas are eliminated. Construct a weight calculation model and dynamically assign weights to feature points, using the following formula: For the first The weight values ​​for each feature point range from 0.1 to 0.

9. For the first The gray-level variance value of each feature point For the first Gradient magnitude values ​​at each feature point; These are the weighting coefficients related to the gray-level variance of the feature points. These are the weighting coefficients for the gradient magnitude at the feature points; For the first The neighborhood consistency coefficient of each feature point ranges from 0 to 1. Dynamic matching of multiple templates: Three types of basic rail profile templates are preset: standard rail, lightly worn rail, and turnout area rail. Based on the feature points after dynamic weight allocation, the hausdorf distance between the inner profile data of the rail and the three types of basic rail profile templates is calculated and optimized, and the template with the highest similarity is automatically selected. A local iterative matching strategy is adopted, with a fixed iteration step size, to continuously optimize the rotation matrix and translation vector of the adaptation template, and finally output the optimal rotation and translation matrix of the standard contour and the measured contour.

3. The orbit detection system based on adaptive matching and multi-source fusion according to claim 1, characterized in that, The specific working process of the federal Kalman filter fusion module is as follows: Establish the system state equations and observation equations: The system state vector includes the position, velocity, attitude angle, and sensor zero bias of the carrier; the observation vector includes the motion attitude and velocity data output by the laser inertial navigation unit, the relative mileage data output by the odometer, and the real-time positioning data output by the GNSS positioning module. Dynamically assign data weights based on the scenario: The GNSS positioning module receives GNSS signal strength to identify whether the scene is an open-air scene or a tunnel scene; when it is an open-air scene, scene adaptive weights are assigned to the motion attitude and speed data, the relative mileage data and the real-time positioning data in the observation vector; When it is a tunnel scene, the real-time positioning data is masked, and scene-adaptive weights are assigned to the motion posture and speed data and the relative mileage data. Federal Kalman filter fusion: The system uses a federated Kalman filter algorithm to fuse multi-source data, including motion attitude and velocity data, relative mileage data, and real-time positioning data, and iteratively updates the system state vector to calculate the trajectory of the vehicle in the geodetic coordinate system.

4. The track detection system based on adaptive matching and multi-source fusion according to claim 1, characterized in that, The specific working process of the trajectory mapping module is as follows: The initial pose calibration of the line structured light 3D camera and the laser inertial navigation unit is completed based on the rail profile calibration plate, and a fixed transformation relationship between the camera coordinate system and the inertial navigation coordinate system is established. The fixed transformation relationship includes a rotation matrix and a translation vector. Combining the optimal rotation and translation matrix output by the adaptive weighted profile matching module, the real-time pose of the carrier relative to the left and right rails at each trajectory point is calculated. The real-time pose includes translation amount and rotation angle. Based on the fixed transformation relationship between the camera coordinate system and the inertial navigation coordinate system, the motion trajectory of the carrier in the geodetic coordinate system is transformed to the rail coordinate system, and then mapped to the real trajectory of the midpoint of the left and right rail tops, which serves as the real alignment of the rail.

5. The track detection system based on adaptive matching and multi-source fusion according to claim 1, characterized in that, The specific working process of the sleeper counting module is as follows: The image data of the fastener area is preprocessed, including grayscale conversion, histogram equalization, and noise suppression. A deep learning object detection model is used to identify fasteners in the preprocessed fastener area image, and the position coordinates and corresponding confidence scores of the fasteners are output. Based on the distribution characteristics and spacing consistency of the fasteners, a sliding window is used to calculate the average value of the sleeper spacing within a certain length. After removing duplicate detection results with the same mileage and erroneous detection results with abnormal spacing, a continuous sleeper number is generated starting from the detection starting point. By combining the relative mileage data of the odometer after multi-source data fusion and calibration, the correspondence between track geometric parameters, rail wear data and corresponding sleeper numbers is established, and a detection data association table carrying sleeper numbers is generated and output.

6. The track detection system based on adaptive matching and multi-source fusion according to claim 1, characterized in that, The specific working process of the geometric parameter extraction module is as follows: Based on the actual trajectory of the midpoint of the left and right rail tops, and combined with the midpoint chord measurement method and parameter calculation model, the track geometric parameters are extracted. The track geometric parameters include track orientation, elevation, track gauge, level, superelevation, and triangular pit. Based on the actual trajectory of the midpoints of the left and right rail tops, the rail top coordinates P1 (X1,Y1,Z1), P2 (X2,Y2,Z2), and P3 (X3,Y3,Z3) are extracted; P2 is the midpoint of P1 and P3 along the track mileage direction. Based on P1, P2, and P3, the track direction and elevation parameters are calculated using the midpoint chord measurement method. The horizontal and superelevation parameters are calculated based on the difference in elevation between the left and right rail tops and the rail surface inclination. The formula is as follows: , The standard distance between the midpoints of the left and right rail tops. The roll attitude angle of the carrier; The parameters of the triangular pit are calculated based on the algebraic difference between the horizontal differences of the two cross-sections within the base length. The formula is as follows: , and These are the horizontal measurements of two cross sections within the base length; Tangents are drawn to the vertices of the actual trajectories of the midpoints of the left and right rails respectively. The tangents under the same cross section are shifted downward by a set value. The Euclidean distance between the intersection points of the left and right rail profile data is calculated to obtain the track gauge parameters.

7. The track detection system based on adaptive matching and multi-source fusion according to claim 5, wherein the deep learning target detection model is the YOLOv8 model, and the confidence level of fastener recognition is ≥0.95; the size of the sliding window is adapted to the length of the sleeper.

8. The track detection system based on adaptive matching and multi-source fusion according to claim 1, characterized in that, The system is adapted to various operating environments, including open-air mainline operation scenarios, tunnel operation scenarios, and turnout area operation scenarios.

9. A track detection system based on adaptive matching and multi-source fusion according to any one of claims 1-8, wherein the number of line structured light 3D cameras is 4; the GNSS positioning module provides RTK centimeter-level positioning; and the odometer outputs 1 pulse for every 1 mm traveled, providing a mileage trigger reference for data synchronization and positioning.

10. A trajectory detection method based on adaptive matching and multi-source fusion, employing the trajectory detection system based on adaptive matching and multi-source fusion as described in claim 1, characterized in that, include: The frame body is mounted on the rail to be inspected, the attitude of the line structured light 3D camera is adjusted, the initial pose calibration of the line structured light 3D camera and the laser inertial navigation unit is completed through the rail profile calibration plate, the preset rail profile template and the pre-trained fastener detection deep learning target detection model are loaded, and the parameters are initialized. The synchronization controller receives the coded mileage output by the odometer in real time. When the coded mileage reaches or exceeds the mileage point in the predetermined mileage sequence generated at fixed spatial intervals for the first time, the synchronization controller outputs a trigger pulse, which synchronously triggers the structured light 3D camera to collect the inner profile data of the rail and the image data of the fastener area, the laser inertial navigation unit to collect the motion attitude and speed data of the carrier, and the GNSS positioning module to collect real-time positioning data, thereby realizing the spatiotemporal synchronization of multi-source detection data. The processing computer uses a federated Kalman filter fusion module to fuse laser inertial navigation data, odometry relative mileage data, and GNSS positioning data to calculate the trajectory of the vehicle in the geodetic coordinate system. An adaptive weighted profile matching algorithm is used to sequentially perform anti-interference optimization, feature point extraction, and multi-template dynamic matching on the inner profile data of the rail. The optimal rotation and translation matrix corresponding to the standard rail profile and the measured rail profile is output. Combined with the initial pose calibration results, the real-time pose of the carrier relative to the left and right rails is calculated. At the same time, based on the matching difference between the standard rail profile and the measured rail profile, the vertical wear and side wear of the rail are calculated as rail wear data. Based on the motion trajectory of the carrier in the geodetic coordinate system and the real-time position and posture of the carrier relative to the left and right rails, the motion trajectory of the carrier is transformed into the real trajectory of the midpoint of the top of the left and right rails. Using the midpoint chord measurement method and a preset parameter calculation model, the track orientation, elevation, gauge, level, superelevation, and triangular pit geometric parameters of the track are extracted from the actual trajectory. The image data of the fastener area acquired by the line structured light 3D camera is preprocessed, and the fastener targets in the image are identified by the deep learning target detection model to complete the automatic counting of sleepers and generate continuous sleeper numbers; combined with the relative mileage data of the odometer after multi-source data fusion and calibration, the correspondence between rail wear data, track geometric parameter anomaly data and corresponding sleeper numbers is established. By integrating data on rail wear, abnormal track geometry parameters, and sleeper numbers, a standardized track inspection report is generated, providing data support for track maintenance and fine-tuning operations.

Images