Unmanned aerial vehicle measurement method for dynamic vector change of railroad bridge
By acquiring high-resolution image sequences through drone hovering, establishing a piecewise polynomial model using binarization and edge detection, and eliminating errors by combining perspective transformation, high-precision railway bridge deflection measurement by drone in adverse weather and complex environments was achieved, solving the measurement challenges in existing technologies.
Patent Information
- Application Number
- CN202610063905.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-19
- Publication Date
- 2026-05-12
AI Technical Summary
Existing methods for measuring railway bridge deflection using drones are difficult to achieve high-precision, non-contact measurement in adverse weather and complex geographical environments, and also suffer from attitude correction errors and pixel noise issues.
High-resolution image sequences are acquired by hovering UAVs. A piecewise polynomial model is established through binarization and edge detection. Combined with a perspective transformation model, the dynamic deformation of the bridge is directly measured, eliminating attitude correction errors and pixel noise, thus achieving non-contact measurement.
It achieves high-precision UAV measurement in harsh weather and complex environments, reduces attitude correction errors and single-pixel errors, and achieves a measurement accuracy of 0.02 pixels, which is suitable for dynamic alignment and track irregularity detection of railway bridges.
Smart Images

Figure CN122024095A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of railway bridge deflection measurement, specifically relating to an unmanned aerial vehicle (UAV) method for measuring the dynamic sagittal changes of railway bridges. Background Technology
[0002] By the end of 2024, my country's railway operating mileage reached 162,000 kilometers, including 48,000 kilometers of high-speed rail. Among these, there were over 600,000 simple-supported beam bridges, more than 8,000 continuous beam bridges, and over 140 long-span bridges. In bridge inspections in special areas such as high-altitude, frigid, and mountainous regions, and those spanning large rivers, conventional inspection methods have limitations and blind spots. Especially during inclement weather, routine inspections become even more difficult, resulting in low efficiency, high difficulty, and high risk, while also affecting the accuracy and effectiveness of the inspections. Currently, there are nearly 20,000 simple-supported beam bridges in my country with over 40 years of service, and this number is increasing annually. Given the continuous increase in traffic volume and load, how to quickly inspect and evaluate the service performance of these bridges to ensure safe railway operation is a significant challenge for railway bridge maintenance.
[0003] The static and dynamic alignment of railway bridges are important indicators for evaluating their performance. Current methods for measuring bridge alignment and deflection include displacement gauges, photoelectric deflectometers, microwave radar, total stations, inclinometers, and static levels. However, for bridges spanning rivers, mountains, canyons, or road interchanges, the mid-span location is difficult to reach, making traditional measurement methods challenging and costly. Therefore, research on static and dynamic alignment measurement technologies for railway bridges based on unmanned aerial vehicles (UAVs) and computer vision is urgently needed to provide accurate deformation data for bridge health monitoring and safety performance assessment.
[0004] UAV measurement of the dynamic alignment and deflection of railway cable-stayed bridges presents challenges, including the trade-off between a large field of view and high precision, the problem of UAV swaying, and pixel noise during the measurement process. Currently, UAV methods for measuring bridge deflection mainly include the following: Firstly, the fixed reference point compensation method involves setting a bridge pier as a reference point or a target at the beam end as a reference point to compensate for the UAV's attitude before measuring the deflection at the point to be measured. However, this method requires setting a target on the beam or bridge pier, making it impossible to achieve true targetless remote sensing. Secondly, the laser projection point compensation method. This method compensates for the attitude of the UAV by projecting a fixed laser point from under the bridge onto the bridge as a fixed point. However, the laser point using this method will drift due to air turbulence, atmospheric refraction, ground vibration, etc., introducing new errors. Thirdly, the segmented shooting and stitching method. This method uses a drone or a fixed camera to take close-up photos of the main beam, and then stitches the photos together to obtain a high-resolution image of the main beam. However, this method requires the drone to move while shooting, which cannot meet the high-frequency data acquisition requirements when trains pass by.
[0005] In conclusion, developing a practical method for measuring the deflection of railway bridges using unmanned aerial vehicles (UAVs) remains a critical issue that urgently needs to be addressed in the field of bridge inspection technology. Summary of the Invention
[0006] This invention is proposed to solve the problems existing in the prior art, and its purpose is to provide an unmanned aerial vehicle (UAV) measurement method for dynamic vector changes of railway bridges.
[0007] The technical solution of this invention is: a UAV measurement method for dynamic sag variation of railway bridges, comprising the following steps: A. The drone hovers and collects images, resulting in a high-resolution, uncompressed sequence of images; B. Perform binarization and edge detection on the sequential images to obtain continuous edge contours; C. Establish a piecewise polynomial model with continuity constraints for continuous edge contours to obtain a smooth edge contour geometric expression; D. Based on the smooth edge contour geometry, a pixel displacement sequence characterizing the dynamic deformation of the bridge is obtained; E. Establish a perspective transformation model to obtain the dynamic conversion coefficients between the current frame pixels and the actual length in the image; F. Based on the obtained pixel displacement sequence and dynamic conversion coefficient, the accurate deflection value of the bridge's dynamic vertical displacement under train load is obtained; G. Repeat the above steps to extract deflection as the train passes.
[0008] Furthermore, step A involves the drone hovering and collecting images to obtain a sequence of high-resolution, uncompressed images. The specific process is as follows: First, a high-resolution drone equipped with a stabilization gimbal was selected and hovered at the pre-set optimal observation position on the side of the main beam; Then, high-definition sequential images of the train from entering the bridge to completely leaving the bridge are captured at a specific frame rate per second.
[0009] Furthermore, step B performs binarization and edge detection on the sequence images sequentially to obtain continuous edge contours. The specific process is as follows: First, the obtained image sequence is binarized using Otsu's method; Then, the Canny edge detection algorithm is applied to the binarized image to obtain continuous edge contours through non-maximum suppression and double threshold hysteresis connection.
[0010] Furthermore, step C establishes a piecewise polynomial model with continuity constraints for the continuous edge contour, obtaining a smooth geometric representation of the edge contour. The specific process is as follows: First, a piecewise polynomial model with continuity constraints is established to address the complex deformation characteristics of continuous edge contours. Then, the Sequential Quadratic Programming algorithm is used to iteratively optimize the coefficients of each piecewise polynomial by minimizing the overall fitting error between the actual coordinates of the edge pixels and the calculated coordinates of the model until the convergence condition is met, thereby obtaining a high-precision and smooth geometric representation of the edge contour.
[0011] Furthermore, step D, based on the smooth edge contour geometry, obtains a pixel displacement sequence characterizing the dynamic deformation of the bridge. The specific process is as follows: First, in each frame of the image, based on the fitted piecewise polynomial model, the pixel points corresponding to the fixed supports at both ends of the bridge are determined and connected to form the reference chord. Then, for the selected point to be measured, the image vector value is accurately calculated by solving the pixel coordinate difference between its vertical projection on the polynomial curve of the frame and the reference chord. Finally, the process is executed frame by frame to generate a pixel displacement sequence that characterizes the dynamic deformation of the bridge.
[0012] Furthermore, step E establishes a perspective transformation model to obtain dynamic conversion coefficients between the current frame pixels and the true length in the image. The specific process is as follows: First, to address the perspective projection distortion and slight inter-frame jitter present in drone photography, a perspective transformation model is established in each frame by identifying a pre-defined calibration object with known physical dimensions and utilizing the correspondence between its pixel coordinates and real coordinates. Then, the transformation parameters of the model are solved frame by frame using direct linear transformation; Finally, the dynamic conversion coefficients between pixels and actual length in the current frame are accurately obtained, effectively eliminating projection and jitter errors.
[0013] Furthermore, step F, based on the obtained pixel displacement sequence and dynamic conversion coefficients, yields the precise deflection value of the bridge's dynamic vertical displacement under train load. The specific process is as follows: First, after obtaining the image vector value and dynamic conversion coefficient in the pixel displacement sequence frame by frame, the two are fused together. Then, the instantaneous vector value of the point under test in each frame corresponding to the real physical scale is calculated by scalar multiplication; Next, in order to obtain the final deflection time history curve, the initial frame sequence in which the bridge was in an unloaded state before the train passed was selected, and its average sagittal value was used as the reference. Then, the reference value is subtracted from the image vector values of all subsequent frames, effectively eliminating the influence of the bridge's initial alignment. Finally, the output is the precise deflection value reflecting the dynamic vertical displacement of the bridge under train load.
[0014] Furthermore, step G repeats the above steps, extracting deflection as the train passes, as detailed below: By conducting on-site measurements of time-series images of different trains passing through, steps A-F were repeated to extract the deflection during train passage.
[0015] The beneficial effects of this invention are as follows: This invention eliminates the need for UAV attitude correction. It cleverly uses the main beam's sag as an invariant to decouple the main beam's deflection from the UAV's attitude changes. It fully utilizes the large sample size generated by the high pixel density of machine vision and leverages the physical property of the continuous second derivative of the main beam's deformation to directly measure the main beam's deflection curve. This not only reduces the numerical errors introduced by the attitude correction process but also significantly reduces single-pixel random errors through a piecewise curve model with second-order continuous constraints, achieving an accuracy of 0.02 pixels in real-world bridge measurements in outdoor environments.
[0016] This invention eliminates the need for any operation or target installation on the bridge, as well as drone-based panning and filming, achieving truly non-contact measurement.
[0017] This invention can serve as a component of a UAV inspection subsystem for railway bridge deflection, enabling convenient and rapid identification of abnormalities in the dynamic alignment and track irregularities of railway bridges. This invention improves the automation, intelligence, accuracy, and robustness of UAV inspection for railway bridge deflection, providing a solution for establishing a rapid deflection detection subsystem for railway bridges. Attached Figure Description
[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a binarized image of the sequence of images from the present invention; Figure 3 This is a diagram showing the characteristic edge lines of the main beam of the present invention; Figure 4 This is a fitting diagram of the feature lines of the main beam of the present invention; Figure 5 This is a frame-by-frame fitting diagram of the feature lines in the sequence image of the present invention; Figure 6 The pixel conversion coefficient curve is calculated frame by frame in this invention; Figure 7 This is a dynamic linear curve diagram of the present invention; Figure 8This invention provides the mid-span deflection time history curve, the theoretical deflection time history curve, and the deflection time history curve for the direct tracking method. Detailed Implementation
[0019] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: like Figures 1 to 8 As shown, a UAV measurement method for dynamic sag variation of railway bridges includes the following steps: A. The drone hovers and collects images, resulting in a high-resolution, uncompressed sequence of images; B. Perform binarization and edge detection on the sequential images to obtain continuous edge contours; C. Establish a piecewise polynomial model with continuity constraints for continuous edge contours to obtain a smooth edge contour geometric expression; D. Based on the smooth edge contour geometry, a pixel displacement sequence characterizing the dynamic deformation of the bridge is obtained; E. Establish a perspective transformation model to obtain the dynamic conversion coefficients between the current frame pixels and the actual length in the image; F. Based on the obtained pixel displacement sequence and dynamic conversion coefficient, the accurate deflection value of the bridge's dynamic vertical displacement under train load is obtained; G. Repeat the above steps to extract deflection as the train passes.
[0020] Step A involves hovering the drone to collect high-resolution, uncompressed image sequences. The specific process is as follows: First, a high-resolution drone equipped with a stabilization gimbal was selected and hovered at the pre-set optimal observation position on the side of the main beam; Then, high-definition sequential images of the train from entering the bridge to completely leaving the bridge are captured at a specific frame rate per second.
[0021] Step B involves performing binarization and edge detection on the sequence of images sequentially to obtain continuous edge contours. The specific process is as follows: First, the obtained image sequence is binarized using Otsu's method; Then, the Canny edge detection algorithm is applied to the binarized image to obtain continuous edge contours through non-maximum suppression and double threshold hysteresis connection.
[0022] Step C establishes a piecewise polynomial model with continuity constraints for continuous edge contours to obtain a smooth geometric representation of the edge contours. The specific process is as follows: First, a piecewise polynomial model with continuity constraints is established to address the complex deformation characteristics of continuous edge contours. Then, the Sequential Quadratic Programming algorithm is used to iteratively optimize the coefficients of each piecewise polynomial by minimizing the overall fitting error between the actual coordinates of the edge pixels and the calculated coordinates of the model until the convergence condition is met, thereby obtaining a high-precision and smooth geometric representation of the edge contour.
[0023] Step D, based on the smooth edge contour geometry, obtains the pixel displacement sequence characterizing the dynamic deformation of the bridge. The specific process is as follows: First, in each frame of the image, based on the fitted piecewise polynomial model, the pixel points corresponding to the fixed supports at both ends of the bridge are determined and connected to form the reference chord. Then, for the selected point to be measured, the image vector value is accurately calculated by solving the pixel coordinate difference between its vertical projection on the polynomial curve of the frame and the reference chord. Finally, the process is executed frame by frame to generate a pixel displacement sequence that characterizes the dynamic deformation of the bridge.
[0024] Step E establishes a perspective transformation model to obtain the dynamic conversion coefficients between the current frame pixels and the true length in the image. The specific process is as follows: First, to address the perspective projection distortion and slight inter-frame jitter present in drone photography, a perspective transformation model is established in each frame by identifying a pre-defined calibration object with known physical dimensions and utilizing the correspondence between its pixel coordinates and real coordinates. Then, the transformation parameters of the model are solved frame by frame using direct linear transformation; Finally, the dynamic conversion coefficients between pixels and actual length in the current frame are accurately obtained, effectively eliminating projection and jitter errors.
[0025] Step F, based on the obtained pixel displacement sequence and dynamic conversion coefficients, yields the precise deflection value of the bridge's dynamic vertical displacement under train load. The specific process is as follows: First, after obtaining the image vector value and dynamic conversion coefficient in the pixel displacement sequence frame by frame, the two are fused together. Then, the instantaneous vector value of the point under test in each frame corresponding to the real physical scale is calculated by scalar multiplication; Next, in order to obtain the final deflection time history curve, the initial frame sequence in which the bridge was in an unloaded state before the train passed was selected, and its average sagittal value was used as the reference. Then, the reference value is subtracted from the image vector values of all subsequent frames, effectively eliminating the influence of the bridge's initial alignment. Finally, the output is the precise deflection value reflecting the dynamic vertical displacement of the bridge under train load.
[0026] Step G repeats the above steps, extracting deflection as the train passes. The specific process is as follows: By conducting on-site measurements of time-series images of different trains passing through, steps A-F were repeated to extract the deflection during train passage.
[0027] Specifically, step A involves the drone hovering and collecting images to obtain a sequence of high-resolution, uncompressed images, as detailed below: a1. When recording, use a slight upward angle, with the bottom edge of the beam as the background. When recording a video of a vehicle passing by, make sure that the bottom edge of the beam is in the middle of the field of view as much as possible, and ensure that both ends of the beam are completely visible in the field of view at the same time. a2. The drone camera uses a high bitrate camera to avoid errors caused by storage compression. At the same time, compression should be avoided during export and transmission to prevent resolution loss. a3. The acquired sequence images are as follows: (1); In the formula, k is the image index (k=1, 2, ..., N), u , v () represents the two-dimensional coordinates of image pixels. This is the color vector (RGB) at that pixel location; ,in These represent the intensity values of the red, green, and blue channels of that pixel, respectively.
[0028] Specifically, step B involves performing binarization and edge detection on the sequence images sequentially to obtain continuous edge contours. The specific process is as follows: b1. Calculate the adaptive thresholding method and binarize the image; b11. Convert the acquired sequence of images into grayscale images; (2); In the k In the image, let the grayscale value be... i The number of pixels is n i Then the total number of pixels in the image is N = n 0 +n1+...+n {L-1} .,in L Grayscale level (0~255); b12. Gray level i The probability of occurrence is: (3); b13. Assume the threshold is a The images are divided into two categories: Class C0: Gray level is [0, ...]. a [Background] Class C1: Gray level [ a +1,L -1] refers to the foreground; b14. Calculate the probabilities of the above two classes: (4); b15. Calculate the class mean: (5); b16. Calculate the between-class variance: (6); b17. Define between-class variance as a measure of the degree of separation between two classes: (7); b18. Further derivation yields the following concise formula: (8); b19. Iterate through all possible thresholds a and select the threshold a* with the largest inter-class variance: (9); b110. For the image Binarization is performed using the threshold values described above: (10).
[0029] b2. Use an edge detection algorithm to extract the feature lines of the main beam edge; b21. Convolve the image using a Gaussian kernel to smooth the image and reduce noise. In practice, discrete convolution kernels are more suitable, depending on the kernel size. b × b The weight of each position is: (11); Where i, j = 0, 1, ..., k -1; b22. To maintain the overall brightness of the image, the kernel is normalized: (12); b23. For the image The result of Gaussian blur for: (13); b24. Use the Sobel operator to calculate the gradient of the image in the x and y directions: (14); b25. Calculate image gradients using the Sobel operator: (15); (16); b26. Calculate the gradient magnitude and direction: (17); (18); b27. After obtaining the gradient magnitude and direction, perform a full scan of the image. If the gradient at a pixel location is a local maximum in that direction, consider using it for the next stage; otherwise, set its pixel value to zero. b28. Set two thresholds: a high threshold and a low threshold. The judgment is based on the gradient magnitude of the pixels. If the amplitude is greater than the high threshold, it is considered a strong edge; If the amplitude is between the low threshold and the high threshold, it is considered a weak edge; If the amplitude is below the low threshold, its pixel value is set to zero; Furthermore, weak edges are considered part of the edge if they are connected to "strong edge" pixels; otherwise, they are set to zero. b29. Finally, the main beam feature line, consisting of a series of discrete pixel coordinates, is obtained, i.e., the continuous edge contour: (19).
[0030] Specifically, step C establishes a piecewise polynomial model with continuity constraints for the continuous edge contour to obtain a smooth geometric representation of the edge contour. The specific process is as follows: c1. The piecewise polynomial divides the main beam into n segments, with each segment satisfying second derivative continuity. The function used is as follows: (20); Where j = 0, 1, 2, 3...n, and satisfies the following condition in all internal nodes. C at the location 2 Continuity condition.
[0031] c2. Solve the above optimization problem using sequential quadratic programming to obtain the parameters of the piecewise polynomial; c21. Construct the following objective function: (twenty one); in, β It is a coefficient vector; c22. Construct the Lagrange function: (twenty two); in, λ They are Lagrange multipliers; c23. In the k In the next iteration, at the current point Construct a quadratic programming subproblem: (twenty three); in, It is the search direction; It is the Hessian approximation of the Lagrange function; c24. Solve the above SQP problem to obtain the search direction. d k and new multiplier estimates ; The KKT conditions are as follows: (twenty four); Where, c is along the direction Perform a line search to determine the step size. : (25); Use the following value functions (26); in, This is the penalty parameter.
[0032] Update using BFGS formula : (27); The convergence criteria are set as follows: First-order optimality condition ; Constraint violation ; relative change ; Specifically, step D includes the following equations: (29); In the formula, k For frame number, For the vector curve, For edge fitting curves, For the first k The frame is based on the fitted line endpoints ( x k0 , S k ( x k0 ) (x kn S k (x) mn The line connecting the two chords is the baseline chord.
[0033] Specifically, step E establishes a perspective transformation model to obtain dynamic conversion coefficients between the current frame pixels and the actual length in the image. The specific process is as follows: Let the length of the main beam be L, the coordinates of the left endpoint of the main beam be (x0, S(x0)), and the coordinates of the right endpoint be (x0, S(x0)). n S(x) n Then, the world size represented by a single pixel is... (30).
[0034] Specifically, step F, based on the obtained pixel displacement sequence and dynamic conversion coefficients, yields the precise deflection value of the bridge's dynamic vertical displacement under train load. The specific process is as follows: (31); in, The dynamic alignment and deflection of the main beam in frame m. This represents the average value of the bridge in the first n frames when it was unloaded.
[0035] Example 1 The implementation method of this invention is introduced by taking the dynamic alignment measurement of a steel truss cable-stayed bridge as an example.
[0036] A certain steel truss cable-stayed bridge is a double-tower, double-cable-plane, four-line steel truss cable-stayed bridge with a main span of 480m and a main truss height of 14m.
[0037] A. A DJI Inspire 3 drone was selected and hovered on the side of the main beam to capture high-definition sequential images of the train from entering the bridge to completely leaving the bridge at a speed of 25 frames per second; When recording a video of a vehicle passing by, it is important to ensure that the lower edge of the beam is in the center of the field of view as much as possible, and that both ends of the beam are fully visible in the field of view at the same time. The main beam to be measured should be fully displayed in the center of the screen and should occupy about 4 / 5 of the entire screen width. Recorded in 8K mode, the sequence of images uses CinemaDNG format to avoid compression errors during image storage.
[0038] B. Binarize the image sequence using Otsu's method, such as... Figure 2 As shown, the Canny edge detection algorithm, through non-maximum suppression and double-threshold hysteresis connection, obtains continuous edge contours, laying the foundation for subsequent geometric measurements, such as... Figure 3 As shown.
[0039] C. To address the complex deformation characteristics of bridge edges, a piecewise curve model with continuity constraints is established. The Sequential Quadratic Programming (SQP) algorithm is employed to iteratively optimize the coefficients of each piecewise polynomial by minimizing the overall fitting error between the actual coordinates of edge pixels and the calculated coordinates of the model, until the convergence condition is met. This yields a high-precision and smooth geometric representation. Figure 4 and Figure 5 As shown.
[0040] D. In each frame of the image, based on the fitted piecewise curve model, the pixels corresponding to the fixed supports at both ends of the bridge are first determined and connected to form a reference chord. Then, for selected test points such as the mid-span point, the pixel coordinate difference between its vertical projection on the polynomial curve of that frame and the reference chord is solved to accurately calculate the image vector value. This process is automated frame by frame, ultimately generating a pixel displacement sequence that characterizes the dynamic deformation of the bridge.
[0041] E. To address perspective projection distortion and minor inter-frame jitter in drone photography, a perspective transformation model is established in each frame by identifying pre-defined calibration objects of known physical dimensions or fixed-size components of a bridge, and utilizing the correspondence between their pixel coordinates and real coordinates. A direct linear transformation is then used to solve for the transformation parameters of this model frame by frame, thereby accurately obtaining the dynamic conversion coefficients between pixels and real lengths in the current frame, effectively eliminating projection and jitter errors. Figure 6 As shown; In this example, the actual world size represented by a single pixel calculated frame by frame is 77.02mm~77.10mm.
[0042] F. After obtaining the image vector value and pixel conversion coefficient frame by frame, the two are fused, and the instantaneous vector value of the test point at each frame corresponding to the real physical scale is calculated through scalar multiplication. To obtain the final deflection time history curve, the initial frame sequence in which the bridge is in an unloaded state before the train passes is selected, and its average vector value is used as the reference. The instantaneous vector value of all subsequent frames is subtracted from this reference value. This differential processing effectively eliminates the influence of the initial alignment of the bridge, and the final output is an accurate deflection value reflecting the dynamic vertical displacement of the bridge under train load, such as... Figure 7 and Figure 8 As shown.
[0043] G. By analyzing time-series images of different trains passing through on-site measurements, the deflection during train passage is extracted using steps one through six. Figure 8As can be seen, the direct tracking method used in this example failed to capture the mid-span deflection change of the bridge during train passage due to excessive error. However, the method proposed in this invention accurately captured the main beam deflection change process during train passage. The maximum measured vertical deflection at mid-span during train passage was 72.70 mm, with a standard deviation of 1.32 mm (0.02 pixels). The maximum theoretical deflection at mid-span during train passage was 93.275 mm, and the ratio of the measured to the theoretical deflection was 0.78, consistent with the typical value of 0.7 to 0.8 for the verification coefficient in load tests.
[0044] Thus, the UAV measurement method for railway bridges based on vector variation is completed.
Claims
1. A UAV measurement method for dynamic vector variation of railway bridges, characterized in that: Includes the following steps: A. The drone hovers and collects images, resulting in a high-resolution, uncompressed sequence of images; B. Perform binarization and edge detection on the sequential images to obtain continuous edge contours; C. Establish a piecewise polynomial model with continuity constraints for continuous edge contours to obtain a smooth edge contour geometric expression; D. Based on the smooth edge contour geometry, a pixel displacement sequence characterizing the dynamic deformation of the bridge is obtained; E. Establish a perspective transformation model to obtain the dynamic conversion coefficients between the current frame pixels and the actual length in the image; F. Based on the obtained pixel displacement sequence and dynamic conversion coefficient, the accurate deflection value of the bridge's dynamic vertical displacement under train load is obtained; G. Repeat the above steps to extract deflection as the train passes.
2. The UAV measurement method for dynamic vector variation of railway bridges according to claim 1, characterized in that: Step A involves hovering the drone to collect high-resolution, uncompressed image sequences. The specific process is as follows: First, a high-resolution drone equipped with a stabilization gimbal was selected and hovered at the pre-set optimal observation position on the side of the main beam; Then, high-definition sequential images of the train from entering the bridge to completely leaving the bridge are captured at a specific frame rate per second.
3. The UAV measurement method for dynamic vector variation of railway bridges according to claim 1, characterized in that: Step B involves performing binarization and edge detection on the sequence of images sequentially to obtain continuous edge contours. The specific process is as follows: First, the obtained image sequence is binarized using Otsu's method; Then, the Canny edge detection algorithm is applied to the binarized image to obtain continuous edge contours through non-maximum suppression and double threshold hysteresis connection.
4. The UAV measurement method for dynamic vector variation of railway bridges according to claim 1, characterized in that: Step C establishes a piecewise polynomial model with continuity constraints for continuous edge contours to obtain a smooth geometric representation of the edge contours. The specific process is as follows: First, a piecewise polynomial model with continuity constraints is established to address the complex deformation characteristics of continuous edge contours. Then, the Sequential Quadratic Programming algorithm is used to iteratively optimize the coefficients of each piecewise polynomial by minimizing the overall fitting error between the actual coordinates of the edge pixels and the calculated coordinates of the model until the convergence condition is met, thereby obtaining a high-precision and smooth geometric representation of the edge contour.
5. The UAV measurement method for dynamic vector variation of railway bridges according to claim 1, characterized in that: Step D, based on the smooth edge contour geometry, obtains the pixel displacement sequence characterizing the dynamic deformation of the bridge. The specific process is as follows: First, in each frame of the image, based on the fitted piecewise polynomial model, the pixel points corresponding to the fixed supports at both ends of the bridge are determined and connected to form the reference chord. Then, for the selected point to be measured, the image vector value is accurately calculated by solving the pixel coordinate difference between its vertical projection on the polynomial curve of the frame and the reference chord. Finally, the process is executed frame by frame to generate a pixel displacement sequence that characterizes the dynamic deformation of the bridge.
6. The UAV measurement method for dynamic vector variation of railway bridges according to claim 1, characterized in that: Step E establishes a perspective transformation model to obtain the dynamic conversion coefficients between the current frame pixels and the true length in the image. The specific process is as follows: First, to address the perspective projection distortion and slight inter-frame jitter present in drone photography, a perspective transformation model is established in each frame by identifying a pre-defined calibration object with known physical dimensions and utilizing the correspondence between its pixel coordinates and real coordinates. Then, the transformation parameters of the model are solved frame by frame using direct linear transformation; Finally, the dynamic conversion coefficients between pixels and actual length in the current frame are accurately obtained, effectively eliminating projection and jitter errors.
7. The UAV measurement method for dynamic vector variation of railway bridges according to claim 1, characterized in that: Step F, based on the obtained pixel displacement sequence and dynamic conversion coefficients, yields the precise deflection value of the bridge's dynamic vertical displacement under train load. The specific process is as follows: First, after obtaining the image vector value and dynamic conversion coefficient in the pixel displacement sequence frame by frame, the two are fused together. Then, the instantaneous vector value of the point under test in each frame corresponding to the real physical scale is calculated by scalar multiplication; Next, in order to obtain the final deflection time history curve, the initial frame sequence in which the bridge was in an unloaded state before the train passed was selected, and its average sagittal value was used as the reference. Then, the reference value is subtracted from the image vector values of all subsequent frames, effectively eliminating the influence of the bridge's initial alignment. Finally, the output is the precise deflection value reflecting the dynamic vertical displacement of the bridge under train load.
8. The UAV measurement method for dynamic vector variation of railway bridges according to claim 1, characterized in that: Step G repeats the above steps, extracting deflection as the train passes. The specific process is as follows: By using on-site measurements of time-series images of different trains passing through, steps A to F are repeated to extract the deflection of the trains as they pass through.
