Wheel-rail attack angle detection method based on machine learning and binocular vision

Abstract

The present application belongs to the field of rail transit safety and the field of intelligent operation and maintenance of rail vehicles, and relates to a wheel-rail attack angle detection method based on machine learning and binocular vision. The method continuously acquires high-quality image sequences of the wheel-rail contact area under complex working conditions by deploying a hardware acquisition system integrated with a binocular camera, a line laser, and a light supplement device at the end of the vehicle bogie. The system internally constructs a visual detection architecture of "interest region pre-screening-key point positioning-three-dimensional coordinate transformation", and the core driving force is a top-down feature point detection network introducing a global attention mechanism. In the actual measurement operation stage, the synchronously collected video stream is input into the network, and the two-dimensional coordinates of each of the two virtual feature points projected on the left and right sides of the single track surface are accurately positioned frame by frame; then, relying on the binocular parallax principle and the three-dimensional coordinate transformation matrix, the above pixel coordinates are restored to the real three-dimensional physical space coordinates.

Description

Technical Field

[0001] This invention belongs to the fields of rail transit safety and intelligent operation and maintenance of rail vehicles, and particularly relates to a wheel-rail angle of attack detection method based on machine learning and binocular vision. Background Technology

[0002] Train safety and equipment stability are core requirements for rail transit operation. During train operation, the contact state between the wheels and rails directly affects vehicle operating safety and equipment lifespan. Among these parameters, the wheel-rail angle of attack is a key parameter reflecting the relative contact posture between the wheels and rails, directly determining the stress and wear on the wheels and rails, and is an important indicator for monitoring the condition of rail transit.

[0003] Wheel-rail angle of attack refers to the angle between the wheel's rolling plane and the longitudinal centerline of the rail. When a train travels at high speed or traverses curved sections, the wheel-rail angle of attack changes dynamically. Excessive angle of attack or abnormal fluctuations can cause uneven wear and surface damage to the wheels and rails, accelerate fatigue aging of the vehicle's running gear components, increase maintenance costs, and in severe cases, lead to abnormal wheel-rail contact and safety hazards. Therefore, real-time and accurate monitoring of the wheel-rail angle of attack is crucial for equipment maintenance, operational optimization, and safe operation.

[0004] However, under complex actual operating conditions, the wheel-rail contact attitude is constantly changing due to multiple factors such as high-speed train operation, track irregularities, real-time and variable wheel-rail contact status, and nonlinear response of the vehicle suspension system. This makes high-precision, real-time online monitoring of the wheel-rail angle of attack extremely difficult, and has always been a technical challenge in the field of rail transit engineering. Currently, the industry lacks direct and effective real-time monitoring methods for the angle of attack. Traditional technologies cannot accurately capture dynamic fluctuation data of the angle of attack during train operation, making it difficult to support wheel-rail wear assessment and refined equipment operation and maintenance.

[0005] From a dynamic transmission perspective, subtle changes in wheel-rail contact first affect the wheelset. As the signal is transmitted from the wheelset to the bogie frame, it is affected by factors such as track disturbances, structural deformation, and suspension damping attenuation, leading to signal attenuation, distortion, and noise superposition. Therefore, the bogie acceleration data cannot accurately reflect the actual wheel-rail contact state and real-time changes in the angle of attack.

[0006] Traditional indirect monitoring methods can only roughly determine macroscopic anomalies in vehicles and cannot accurately capture the microscopic contact state of wheels and rails, making it difficult to meet the needs of accurate wear assessment, preventive maintenance, and refined train operation control. To address this technical limitation, this patent proposes a wheel-rail angle of attack calculation method based on a virtual point tracking network. This method overcomes the limitations of traditional monitoring methods, achieving high-precision real-time calculation of wheel-rail angle of attack, and providing a reliable new technological solution for accurate wheel-rail condition monitoring and safe operation and maintenance of rail vehicles. Summary of the Invention

[0007] To address the shortcomings of existing indirect measurement and two-dimensional monocular vision techniques, this invention combines binocular vision 3D reconstruction technology with a deep learning-based visual detection model based on a top-down architecture. It provides a dynamic 3D wheel-rail angle of attack detection method based on deep learning and binocular vision. The specific steps are as follows: S1. Construct a wheel-rail dual-level visual detection network based on a top-down architecture. The dual-level visual detection network includes a region of interest (ROI) target detection model for pre-screening and a key point detection model for precise localization. S2. A fixture integrating a line laser, supplementary lighting device, and high-definition binocular camera is installed at the end of the bogie. Two displacement sensors are symmetrically installed on the left and right sides of the bogie frame to detect the longitudinal displacement of the wheelset relative to the frame in real time and to collect binocular video of the track during the dynamic driving process of the vehicle. Then, a video segment is selected, and image frames are extracted from the binocular video. The training set and internal validation set are divided in a 7:3 ratio. First, rectangular boxes are used to mark the feature targets containing the wheel-rail contact area in the image to construct the ROI detection dataset. The feature targets are denoted as the detection boxes of a single track. Then, the constructed ROI detection dataset is used to train the untrained target detection model. The loss function adopts a weighted combination of cross-entropy loss and intersection-over-union (IoU) loss of bounding box regression. The LinearLR strategy is used in the warm-up stage, and then the cosine annealing strategy is used to decay the learning rate. After training, the trained ROI target detection model is obtained. S3. Mark key points in the cropped ROI image region to construct a key point detection dataset. Specifically, mark four virtual key points on the surface features of a single track: the first feature point P1 (top left) and the third feature point P3 (bottom left) on the left side of the track, and the second feature point P2 (top right) and the fourth feature point P4 (bottom right) on the right side of the track. Then, use the SimCC classification algorithm to decouple and predict the horizontal and vertical coordinates of the key points. S4. Apply data augmentation strategies, including random affine transformation, HSV color space perturbation, and local occlusion, to the ROI detection dataset and keypoint detection dataset. Then, input the augmented data into the untrained keypoint detection model for training. The optimizer is AdamW, and the loss is calculated using KL divergence (KLDiscretLoss). After warm-up and cosine annealing learning rate scheduling, the trained keypoint detection model is obtained. S5. Enter the actual test operation stage. Input another segment of wheel-rail contact binocular video to be detected, and simultaneously read frame-by-frame displacement data collected by two displacement sensors aligned with the video frame timestamps. Extract the left and right eye images frame by frame and input them into the trained ROI target detection model respectively. Output the predicted target boxes for each binocular image. Crop the left and right eye images according to the predicted target boxes, and uniformly scale the cropped local images to a fixed resolution size (640×640). Input the preprocessed left and right eye local images into the trained key point detection model to predict the pixel coordinates of the key points in the entire image (3840×1080) in each frame; obtain the first... In the frame, the two-dimensional pixel coordinates of the first to fourth feature points in the left-eye full-image coordinate system. And the two-dimensional pixel coordinates in the right-eye full-image coordinate system. ,in =1,2,3,4; S6. Based on the principle of binocular parallax, calibrate the intrinsic and extrinsic parameters of the binocular camera, thereby mapping the two-dimensional pixel coordinates extracted in step S5 to three-dimensional physical space coordinates; first calculate the... Frame 1 Binocular parallax at feature points ;

[0008] Combined with binocular baseline distance Solve for its three-dimensional physical space coordinate matrix in the right eye coordinate system.

[0009] ;

[0010] In the formula, and These are the principal point coordinates of the left and right cameras, respectively. The X-axis focal lengths of the left and right cameras are... The inverse rotation extrinsic parameter matrix of the right camera. This is the inverse intrinsic parameter matrix of the right camera; S7. In a three-dimensional physical space coordinate system, the four feature points form a dynamic geometric parallelogram on the space plane; extract the first feature point on the same side of the track. With the third feature point The connecting line serves as the reference direction line characterizing the current orbital geometry; the first feature point spanning the opposite sides of the orbit... With the second feature point The connecting line is established as the upper angle of attack characterization line, and the third feature point is... With the fourth feature point The connecting line is established as the lower angle of attack characterization line; based on the three-dimensional spatial coordinates of each feature point, the first... Instantaneous slopes of the lower reference direction line, the upper angle of attack characterization line, and the lower angle of attack characterization line. : ; ;

[0011] Subsequently, the arctangent formula was used to calculate the first... Below the frame, the instantaneous spatial angle between the upper angle of attack characterization line and the reference direction line. And the instantaneous spatial angle between the lower angle of attack characterization line and the reference direction line. : ;

[0012] First frame of the video Extracted and As the initial reference angle in the unbiased state, the first angle is obtained through difference calculation. The frame's real-time dynamic deflection angle relative to the rail: ;

[0013] S8. Align the displacement data collected by the two displacement sensors with the video frame timestamps, and read the [timestamp]. Displacement readings of the left and right displacement sensors in the frame and Based on the lateral installation spacing between the two displacement sensors Solve the first The deflection angle of the wheelset relative to the frame in the next frame: ;

[0014] In the formula, For the first The deflection angle of the wheelset relative to the frame in the next frame. , The left and right displacement sensors are respectively in the first... Displacement readings in frames, The lateral mounting spacing between the two displacement sensors; Since the binocular camera is fixed to the bogie frame via tooling, the solution obtained in step S7 is... and Essentially, it is the deflection angle of the frame relative to the rail; by subtracting the deflection angle of the frame relative to the rail from the deflection angle of the wheelset relative to the frame, the coupling effect of the frame's own deflection is eliminated, and the first... The actual wheel-rail angle of attack of the wheelset relative to the rail in the next frame: ;

[0015] In the formula, The first The two sets of actual wheel-rail angles of attack obtained by solving the upper and lower angle of attack characterization lines in the frame are all signed quantities and the positive direction of deflection is consistent. The upper and lower sets of angle of attack data, containing all test video frames, are arranged into vector sets in time sequence. Zero-meaning and smoothing filtering are performed on the two vector sets respectively. Finally, two high-fidelity comparative data curves reflecting the time-domain evolution characteristics of wheel-rail angle of attack are output, and the two curves verify each other.

[0016] This patent aims to achieve direct detection of three-dimensional dynamic wheel-rail angle of attack. Unlike traditional indirect calculations or two-dimensional pixel difference measurements, this patent uses deep learning combined with binocular vision to directly observe key feature points in the wheel-rail contact area and accurately reconstructs them into true three-dimensional physical space angles based on the principle of binocular parallax. This method of directly observing spatial geometric attitude can more intuitively and clearly reflect the true dynamic wheel-rail contact relationship, thus providing direct and reliable data support and early warning for preventing and solving a series of core train safety issues such as train derailment on curves, abnormal and rapid wheel-rail wear, and snake-like instability.

[0017] (1) In the prior art, the traditional indirect measurement method using frame or axle box acceleration sensors is easily affected by the attenuation of the vehicle suspension system, and cannot accurately restore the true wheel-rail spatial geometric attitude. As for the technical means of deploying laser or visual monitoring equipment by the trackside, not only is the equipment expensive and the installation and maintenance cumbersome, but it can only achieve local fixed-point detection. The present invention adopts a vehicle-mounted binocular camera tooling solution, which does not require complex trackside infrastructure, has low equipment cost and is easy to integrate, and can achieve continuous online monitoring of wheel-rail angle of attack for the entire track section with the vehicle.

[0018] (2) This invention breaks away from the limitations of traditional two-dimensional image pixel differences. Relying on the principle of binocular parallax and three-dimensional physical space matrix mapping, it transforms the two-dimensional precise positioning points of each frame into a three-dimensional absolute physical coordinate system for geometric calculation. This three-dimensional ranging and angle measurement method is not affected by image scaling and can accurately capture the tiny spatial deflection between the wheel and the rail, greatly improving the detection accuracy of dynamic angle of attack.

[0019] (3) In terms of network structure, the present invention adopts a top-down two-level visual detection architecture, that is, firstly, the overall wheel-rail contact area is obtained through ROI target detection, and then the local virtual key points are accurately located. This prediction method from the whole to the part can not only effectively filter the background environmental interference under complex working conditions, but also better retain the high-resolution feature information of the target area, significantly improving the positioning reliability and robustness under high-frequency vibration and complex reflective environment.

[0020] (4) The present invention further introduces displacement sensor groups symmetrically installed on the left and right sides of the bogie frame to directly detect the deflection angle of the wheelset relative to the frame and fuse it with the deflection angle of the frame relative to the rail obtained by binocular vision calculation. This effectively eliminates the coupling interference of the frame's own deflection on the measurement results and ensures that the final output is the true angle of attack of the wheelset relative to the rail, further improving the physical authenticity and engineering credibility of the angle of attack detection. Attached Figure Description

[0021] Figure 1 For technical flowcharts; Figure 2 This is a schematic diagram showing the changes in the deflection angle of the frame and the rails; Figure 3 This is a schematic diagram of a track-based laser. Figure 4 This is a schematic diagram of the top-down method detection; Figure 5 This is a schematic diagram for calculating the frame and wheel deflection angles; Figure 6 A comparison chart of the angle of attack and curvature detected using this method on a certain subway line; Reference numerals: 1—Displacement sensor; 2—Axle box; 3—Wheelset; 4—Rail; 5—Frame. Detailed Implementation

[0022] like Figure 1 The specific steps are as follows: S1. Construct a wheel-rail dual-level visual detection network based on a top-down architecture. The dual-level visual detection network includes a region of interest (ROI) target detection model for pre-screening and a key point detection model for precise localization. S2. A fixture integrating a line laser, supplementary lighting device, and high-definition binocular camera is installed at the end of the bogie. Two displacement sensors are symmetrically installed on the left and right sides of the bogie frame to detect the longitudinal displacement of the wheelset relative to the frame in real time and to collect binocular video of the track during the dynamic driving process of the vehicle. Then, a video segment is selected, and image frames are extracted from the binocular video. The training set and internal validation set are divided in a 7:3 ratio. First, rectangular boxes are used to mark the feature targets containing the wheel-rail contact area in the image to construct the ROI detection dataset. The feature targets are denoted as the detection boxes of a single track. Then, the constructed ROI detection dataset is used to train the untrained target detection model. The loss function adopts a weighted combination of cross-entropy loss and bounding box regression IoU loss. The LinearLR strategy is used in the warm-up stage, and then the cosine annealing strategy is used to decay the learning rate. After training, the trained ROI target detection model is obtained. S3. Mark key points in the cropped ROI image region to construct a key point detection dataset. Specifically, mark four virtual key points of the single track surface features: the first feature point P1 (top left) and the third feature point P3 (bottom left) on the left side of the track, and the second feature point P2 (top right) and the fourth feature point P4 (bottom right) on the right side of the track. Then, use the SimCC classification algorithm to decouple and predict the horizontal and vertical coordinates of the key points. S4. Apply data augmentation strategies, including random affine transformation, HSV color space perturbation, and local occlusion, to the ROI detection dataset and keypoint detection dataset. Then, input the augmented data into the untrained keypoint detection model for training. The optimizer is AdamW, and the loss is calculated using KL divergence (KLDiscretLoss). After warm-up and cosine annealing learning rate scheduling, the trained keypoint detection model is obtained. S5. Enter the actual test operation stage. Input another segment of wheel-rail contact binocular video to be detected, and simultaneously read frame-by-frame displacement data collected by two displacement sensors aligned with the video frame timestamps. Extract the left and right eye images frame by frame and input them into the trained ROI target detection model respectively. Output the predicted target boxes for each binocular image. Crop the left and right eye images according to the predicted target boxes, and uniformly scale the cropped local images to a fixed resolution size (640×640). Input the preprocessed left and right eye local images into the trained key point detection model to predict the pixel coordinates of the key points in the entire image (3840×1080) in each frame; obtain the first... In the frame, the two-dimensional pixel coordinates of the first to fourth feature points in the left-eye full-image coordinate system. And the two-dimensional pixel coordinates in the right-eye full-image coordinate system. ,in =1,2,3,4; S6. Based on the principle of binocular parallax, the intrinsic and extrinsic parameters of the binocular camera obtained through calibration are determined using ZED Calibration software, thereby mapping the two-dimensional pixel coordinates extracted in step S5 to three-dimensional physical space coordinates; firstly, the first... Frame 1 Binocular parallax at feature points ;

[0023] Combined with binocular baseline distance Solve for its three-dimensional physical space coordinate matrix in the right eye coordinate system. : ;

[0024] In the formula, These are the principal point coordinates of the left and right cameras, respectively. The X-axis focal lengths of the left and right cameras are... The inverse rotation extrinsic parameter matrix of the right camera. This is the inverse intrinsic parameter matrix of the right camera; S7. In a three-dimensional physical space coordinate system, the four feature points form a dynamic geometric parallelogram on the space plane; extract the first feature point on the same side of the track. With the third feature point The connecting line serves as the reference direction line characterizing the current orbital geometry; the first feature point spanning the opposite sides of the orbit... With the second feature point The connecting line is established as the upper angle of attack characterization line, and the third feature point is... With the fourth feature point The connecting line is established as the lower angle of attack characterization line; based on the three-dimensional spatial coordinates of each feature point, the first... Instantaneous slopes of the lower reference direction line, the upper angle of attack characterization line, and the lower angle of attack characterization line. : ;

[0025] Subsequently, the arctangent formula was used to calculate the first... Below the frame, the instantaneous spatial angle between the upper angle of attack characterization line and the reference direction line. And the instantaneous spatial angle between the lower angle of attack characterization line and the reference direction line. : ; like Figure 2 The first frame of the video Extracted As the initial reference angle in the unbiased state, the first angle is obtained through difference calculation. The frame's real-time dynamic deflection angle relative to the rail: ; S8, such as Figure 3 , Figure 4 Align the displacement data collected by the two displacement sensors with the timestamps of the video frames, and read the timestamps of the video frames. Displacement readings of the left and right displacement sensors in the frame Based on the lateral installation spacing between the two displacement sensors ,like Figure 5 Solve the first The deflection angle of the wheelset relative to the frame in the next frame:

[0026] In the formula, For the first The deflection angle of the wheelset relative to the frame in the next frame. The left and right displacement sensors are respectively in the first... Displacement readings in frames, The lateral mounting spacing between the two displacement sensors; Since the binocular camera is fixed to the bogie frame via tooling, the solution obtained in step S7 is... and Essentially, it is the deflection angle of the frame relative to the rail; by subtracting the deflection angle of the frame relative to the rail from the deflection angle of the wheelset relative to the frame, the coupling effect of the frame's own deflection is eliminated, and the first... The actual wheel-rail angle of attack of the wheelset relative to the rail in the next frame:

[0027] In the formula, The first The two sets of actual wheel-rail angles of attack obtained by solving the upper and lower angle of attack characterization lines in the frame are all signed quantities and the positive direction of deflection is consistent. The upper and lower sets of attack angle data, containing all test video frames, are arranged into vector sets in time sequence. Zero-mean reduction and smoothing filtering are then performed on both vector sets respectively. The final output consists of two high-fidelity comparative data curves reflecting the time-domain evolution characteristics of the wheel-rail attack angle. These two curves are used to verify each other. Figure 6 .

[0028] In a three-dimensional physical coordinate system, the four feature points are connected by a closed line to form a geometric parallelogram. The side connecting two feature points on the same side of the track (e.g., the left side of the track) is established as the reference direction line characterizing the current track geometry. The upper and lower sides, spanning the left and right sides of the track and connecting feature points on opposite sides, serve as the characterizing lines for the deflection angle between the frame and the track. When the vehicle generates a wheel-rail angle of attack, the upper and lower sides undergo instantaneous slope changes with the deflection of the frame relative to the rail. By solving the real-time spatial geometric angle between the track geometry reference direction line and the upper and lower sides, and calculating the difference between this angle and the initial angle under theoretical no-deflection conditions, the real-time dynamic deflection angles of the two sets of frames relative to the rail are calculated. Furthermore, by using two displacement sensors symmetrically installed on the left and right sides of the bogie frame to detect the longitudinal displacement of the wheelset relative to the frame in real time, the deflection angle of the wheelset relative to the frame is calculated, and the difference is made with the deflection angle of the frame relative to the rail to eliminate the coupling effect of the frame's own deflection, thus obtaining the true wheel-rail angle of attack of the wheelset relative to the rail. Finally, the two sets of angle of attack vectors generated are subjected to zero-mean reduction and smoothing filtering, and two data curves that can reflect the time-domain evolution characteristics of wheel-rail angle of attack with high fidelity are finally output.

[0029] The electronic device for wheel-rail angle of attack detection provided by this invention includes at least one high-performance processor and a corresponding data storage unit at the hardware level. The storage unit is configured with a non-volatile or volatile access module, which is used not only to store the program instruction set and algorithm logic for executing the method of this invention, but also to cache wheel-rail image matrices, geometric feature parameters, and angle of attack timing data in real time. The processor, as the core computing hub of the system, is deeply coupled to the storage unit through a bus architecture, enabling it to execute complex computational tasks in parallel, including preprocessing algorithms, two-level visual detection logic, and three-dimensional geometric solution formulas. The processor can be implemented using one or more architectures of a digital signal processor, a field-programmable gate array (FPGA), or a high-performance central processing unit (CPU) to meet the real-time analysis requirements of high-frequency dynamic video streams; optionally, the system can also integrate a communication interface chip for synchronous access to external sensor data and remote transmission of detection results.

[0030] This invention also includes a computer-readable storage medium on which computer-executable instructions are loaded for implementing the three-dimensional wheel-rail angle of attack dynamic detection method of this invention. When the instructions are loaded and invoked by the processing core of an electronic device, the system can be forced to complete the complete technical logic from video data acquisition, automatic tracking of the region of interest, binocular parallax matching, three-dimensional spatial matrix projection, to real-time angle of attack calculation and fluctuation curve output. This medium can be in the form of a magnetic storage medium, optical storage medium, or solid-state storage such as flash memory, aiming to ensure the standardization and repeatability of the wheel-rail angle of attack detection process through programmed execution logic.

Claims

1. A wheel-rail attack angle detection method based on machine learning and binocular vision, characterized in that, Includes the following steps: S1. Construct a wheel-rail dual-level visual detection network based on a top-down architecture. The dual-level visual detection network includes a region of interest target detection model for pre-screening and a key point detection model for precise localization. S2. The tooling integrating the line laser, supplementary lighting device, and high-definition binocular camera is installed at the end of the bogie. Displacement sensors are symmetrically installed on the left and right sides of the bogie frame to detect the longitudinal displacement of the wheelset relative to the frame in real time and to collect binocular video of the track during the dynamic driving process of the vehicle. Then, a binocular video segment is selected from the video, and image frames are extracted from it. The training set and internal validation set are divided proportionally. First, rectangular boxes are used to mark the feature targets containing the wheel-rail contact area in the image to construct the region of interest detection dataset. The feature targets are denoted as the detection boxes of a single track. Then, the constructed region of interest detection dataset is used to train the untrained target detection model. The loss function is a weighted combination of cross-entropy loss and intersection-union ratio loss of bounding box regression. The learning rate is decayed using a cosine annealing strategy. After training, the trained region of interest target detection model is obtained. S3. Mark key points in the cropped region of interest image to construct a key point detection dataset. Specifically, mark four virtual key points of the surface features of a single track: the first feature point P1 and the third feature point P3 on the left side of the track, and the second feature point P2 and the fourth feature point P4 on the right side of the track; and decouple and predict the horizontal and vertical coordinates of the key points. S4. Apply data augmentation strategies, including random affine transformation, color space perturbation, and local occlusion, to the region of interest detection dataset and the keypoint detection dataset. Then, input the augmented data into the untrained keypoint detection model for training. After warm-up and cosine annealing learning rate scheduling, the trained keypoint detection model is obtained. S5. Enter the actual test operation phase. Input another segment of wheel-rail contact binocular video to be detected, and simultaneously read frame-by-frame displacement data collected by two displacement sensors, aligned with the video frame timestamps. Extract the left and right eye images frame by frame and input them into the trained region of interest target detection model respectively. Output the predicted target boxes for each binocular image. Crop the left and right eye images according to the predicted target boxes, and uniformly scale the cropped local images to a fixed resolution. Input the preprocessed left and right eye local images into the trained keypoint detection model to predict the pixel coordinates of keypoints in the entire image for each frame. Obtain the first... In the frame, the two-dimensional pixel coordinates of the first to fourth feature points in the left-eye full-image coordinate system. And the two-dimensional pixel coordinates in the right-eye full-image coordinate system. ,in =1,2,3,4; S6. Based on the principle of binocular parallax, calibrate the intrinsic and extrinsic parameters of the binocular camera obtained from the calibration, thereby mapping the two-dimensional pixel coordinates extracted in step S5 to three-dimensional physical space coordinates; first calculate the... Frame 1 Binocular parallax of feature points: ； Combined with binocular baseline distance Solve for its three-dimensional physical space coordinate matrix in the right eye coordinate system. : ; In the formula, and These are the principal point coordinates of the left and right cameras, respectively. The X-axis focal lengths of the left and right cameras are... The inverse rotation extrinsic parameter matrix of the right camera. This is the inverse intrinsic parameter matrix of the right camera; S7. In a three-dimensional physical space coordinate system, the four feature points form a dynamic geometric parallelogram on the space plane; Extract the first feature point on the same side of the track. With the third feature point The connecting line serves as the reference direction line characterizing the current orbital geometry. The first feature point spanning the left and right sides of the track With the second feature point The connecting line is established as the upper angle of attack characterization line, and the third feature point is... With the fourth feature point The connecting line is established as the lower angle of attack characterization line; based on the three-dimensional spatial coordinates of each feature point, the first... Instantaneous slopes of the lower reference direction line, the upper angle of attack characterization line, and the lower angle of attack characterization line. , and : ; Subsequently, the arctangent formula was used to calculate the first... Below the frame, the instantaneous spatial angle between the upper angle of attack characterization line and the reference direction line, and the instantaneous spatial angle between the lower angle of attack characterization line and the reference direction line. : ; First frame of the video Extracted and As the initial reference angle in the unbiased state, the first angle is obtained through difference calculation. The frame's structure compared to the real-time dynamic deflection angle of the rail: ; S8. Align the displacement data collected by the two displacement sensors with the video frame timestamps, and read the [timestamp]. Displacement readings of the left and right displacement sensors in the frame and Based on the lateral installation spacing between the two displacement sensors Solve the first The deflection angle of the wheel pair relative to the frame: ; In the formula, For the first The deflection angle of the wheelset relative to the frame in the next frame. , The left and right displacement sensors are respectively in the first... Displacement readings in frames, The lateral mounting spacing between the two displacement sensors; Since the binocular camera is fixed to the bogie frame via tooling, the solution obtained in step S7 is... and Essentially, it is the deflection angle of the frame relative to the rails; The deflection angle of the frame relative to the rail is subtracted from the deflection angle of the wheelset relative to the frame to eliminate the coupling effect of the frame's own deflection, thus obtaining the first... The actual wheel-rail angle of attack of the wheelset compared to the rail in the frame: ; ; In the formula, , The first The two sets of actual wheel-rail angles of attack obtained by solving the upper and lower angle of attack characterization lines in the frame are all signed quantities and the positive direction of deflection is consistent. The upper and lower sets of angle of attack data, containing all test video frames, are arranged into vector sets in time sequence. Zero-meaning and smoothing filtering are performed on the two vector sets respectively. Finally, two high-fidelity comparative data curves reflecting the time-domain evolution characteristics of wheel-rail angle of attack are output, and the two curves verify each other.

2. The method according to claim 1, characterized in that, The specific construction method of the dual-level visual detection network in step S1 is as follows: The region of interest target detection model uses ShuffleNetV2 as the backbone feature extraction module, and the YOLOX architecture is connected to the neck and head structures for bounding box coordinate regression and category prediction; The key point detection model also uses the lightweight ShuffleNetV2 as the backbone module, and its head prediction adopts the RTMCChead structure, and a global attention unit containing position encoding and feature adaptive adjustment mechanism is deeply embedded in the RTMCChead structure to enhance the positional reliability under complex high reflectivity conditions.

3. An electronic device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of claim 1 or 2.

4. A computer-readable medium, characterized in that, The computer-readable medium stores a computer program, wherein the computer program is configured to perform the method of claim 1 or 2 when executed.

Images