Railway bridge dynamic chord measurement method based on unmanned aerial vehicle time series images
By using UAV time-series image acquisition and image processing technology, the problem of measuring dynamic chord values of long-span bridges has been solved, realizing non-contact and accurate dynamic chord value measurement, which is suitable for railway bridge inspection in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA RAILWAY ECONOMIC & PLANNING RES INST
- Filing Date
- 2026-03-02
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies cannot effectively measure the dynamic chord values of long-span railway bridges when trains pass. Traditional equipment cannot reflect the real-time changes in the bridge's alignment and is difficult to install in complex environments.
A method based on UAV time-series images was adopted. Panoramic images of the bridge were acquired by a UAV equipped with a stabilization gimbal. Position filters and scale filters were used to track the beam end areas. Piecewise polynomial fitting was performed by combining OTU binarization and Canny edge extraction algorithm to calculate dynamic chord measurements.
It enables non-contact measurement of dynamic chord values of bridges, is suitable for complex environments, and provides high-frequency, continuous deformation data to support bridge stiffness assessment and operation and maintenance decisions.
Smart Images
Figure CN122157049A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of dynamic chord measurement technology for railway bridges, and in particular to a method for dynamic chord measurement of railway bridges based on UAV time-series images. Background Technology
[0002] The smoothness of long-span railway bridges is crucial for train operation safety. Currently, the "High-Speed Railway Engineering Construction Quality Acceptance Standard" (TB 10752-2018) specifies the track geometry under no-train-load conditions, and defines static chord measurements for chord lengths of 10m, 30m, 60m, and 300m. However, for long-span bridges, the train interacts with the bridge, and the bridge alignment is constantly changing. Current dynamic acceptance testing typically uses vehicle acceleration and track quality index (TQI) as metrics, lacking consideration of the bridge's dynamic chord measurements.
[0003] Measuring the dynamic ride comfort of bridges is extremely difficult. Traditional static levels, due to the lag in fluid flow, can only measure static deformation and cannot reflect the real-time alignment when trains pass. Displacement gauges, photoelectric deflectometers, and microwave radars can only measure displacement at a limited number of points and cannot reflect the continuous alignment of the bridge. Commonly used inertial navigation track trolleys can only measure the static chord values of the track and cannot measure the dynamic chord values. In particular, for bridges spanning rivers, mountains, canyons, and road interchanges, the difficulty in accessing the bridge locations and installing equipment necessitates the development of dynamic ride comfort measurement methods for railway bridges based on UAV time-series images. This would provide accurate deformation data for the dynamic ride comfort detection and safety performance assessment of railway bridges. Summary of the Invention
[0004] The purpose of this invention is to provide a method for measuring dynamic chord values of railway bridges based on UAV time-series images, which solves the problems of existing technologies such as inability to measure dynamic chord values, limited measurement methods, narrow applicable scenarios, and difficulty in balancing measurement accuracy and efficiency.
[0005] To achieve the above objectives, this invention provides a method for measuring the dynamic chord values of railway bridges based on UAV time-series images, comprising the following steps: S1. Acquire high-quality time-series images: Use a drone equipped with a stabilization gimbal to hover at a certain distance from the side of the main beam mid-span and collect time-series panoramic images of the entire process of the train from entering the bridge to completely leaving the bridge at a specific frame rate per second. The time-series panoramic images include the edge line of the main beam and the endpoints on both sides of the main beam. S2. Locking and dynamically updating the beam end region: Using an interactive selection method, the beam end region of the main beam is selected with the two ends of the main beam as the center. The directional gradient histogram features and color features are extracted in the beam end region, and the position filter and scale filter are trained. For each subsequent frame, the position filter is used to detect around the previous frame to obtain a new position. The scale filter is used at the new position to detect the optimal scale. The position filter and scale filter are then updated. S3: Obtaining the clean edge line of the main beam: For each frame of the time-series panoramic image, based on the line connecting the endpoints of the main beam after the update position, a certain range is offset to both sides as the possible region of the beam's feature edge line. After binarizing the possible region of the beam's feature edge line using the OTU method, the Canny algorithm is used to extract the edge line of the main beam, and then... Criterion-based noise reduction; S4. Obtain the true shape of the main beam: Fit the edge line of the main beam using a piecewise polynomial with continuous second derivative; calculate the image pixel conversion coefficient using a fixed beam length as a known dimension; multiply the piecewise polynomial fitting result with the pixel conversion coefficient to obtain the true shape of the main beam; S5. Obtain dynamic chord measurement values: Select the fixed chord length of the true alignment of the main beam as the reference chord. For each measuring point on the true alignment of the main beam, calculate the sagittal value of each measuring point relative to the reference chord to obtain the live load chord measurement value of the main beam. The live load chord measurement value of the main beam is superimposed with the static chord measurement value to obtain the dynamic chord measurement value.
[0006] Preferably, the time-series panoramic image in step S1 includes the edge line of the main beam and the endpoints on both sides of the main beam.
[0007] Preferably, in step S1, the drone records images at an elevation angle of 1° to 3°, and the field of view includes clear beam feature lines or edge lines.
[0008] Preferably, the acquired time-series panoramic images are obtained using the following formula: in, For indexing time-series panoramic images, ; These are the two-dimensional coordinates of pixels in the time-series panoramic image; For the first The RGB color vectors at pixel locations in a time-series panoramic image. ,in, This represents the intensity value of the red channel of the pixel. This represents the intensity value of the green channel of the pixel. This represents the intensity value of the blue channel for a given pixel.
[0009] Preferably, step S2 specifically includes the following steps: S21. Using an interactive selection method, select the beam end areas centered on both ends of the main beam, with the left end area of the main beam as... The right end area of the main beam is ,in, The x-coordinate of the top-left corner of the region. The ordinate of the top-left corner of the region. The width of the region, The height of the region; extract the directional gradient histogram features and color features of the main beam end region respectively, using the following formula: in, Features of the directional gradient histogram; Color characteristics; S22, Training Location Filter and Scale Filter: The objective function for training the location filter is set as minimizing a ridge regression problem, as shown in the formula: in, The weight vector to be learned; For in position The eigenvector at that location; For in position Expected output at; For regularization parameters, Weight vector The square of the L2 norm; The objective function is solved in the Fourier domain, as shown in the formula: in, For the Fourier transform of the filter; Fourier transform of the input features; The Fourier transform of the desired output; For regularization parameters; This indicates element-wise multiplication; , Indicates complex conjugation; Training the scaling filter: Dividing scale intervals: Defining the total number of scales as Using indexes The formula for distinguishing different scales is: ; The width and height of the target region are scaled according to a scale factor to obtain the region size for each scale, as shown in the formula: in, , These are the width and height of the target area, respectively. This is the scale factor, typically 1.02; This is a scale index, indicating the number of scaling points or levels relative to the base scale; Extracting scale features: For the target region at each scale, extract the feature vector, using the following formula: in, Scale features; , , These are the feature vectors at each scale; Define the training objective function: The objective function for training the scaling filter is set as minimizing the ridge regression problem, as shown in the formula: in, The weight coefficients to be learned The desired output is typically a Gaussian-shaped response. For regularization parameters; Weight vector The square of the L2 norm; Solving in the Fourier domain: Transform the objective function to the Fourier domain and solve it. The formula is as follows: in, This represents the scaling filter in the frequency domain. Fourier transform; For the expected output Fourier transform; scale features Fourier transform; This indicates element-wise multiplication; , To obtain the conjugate; For regularization parameters; S23. For each subsequent frame, a position filter is used to detect the area around the previous frame to obtain the new position; the position response is calculated according to the following formula: in, This is the location response result; The Fourier transform of the location feature vectors of the new candidate region; The autocorrelation term representing the model, For autocorrelation terms Fourier transform. Indicates the inverse Fourier transform; For the Fourier transform of the filter, take the complex conjugate; The scale response is calculated using the following formula: in, Fourier transform of the scale feature vectors of the new candidate region; The autocorrelation term represents the model; For the frequency domain representation of the scaling filter, take the complex conjugate; For the autocorrelation term of the model, take the complex conjugate; The updated version The left end region of the main beam of the frame is The right end area of the main beam is ; S24. At the new location, use a scaling filter to detect the optimal scale, and update the location filter and the scaling filter; the location filter update formula is: in, for The filter weights after frame update; for Frame filter weights; for The autocorrelation energy term after frame update; for The autocorrelation energy term of the frame; The learning rate; For time indexing; The scaling filter update formula is: in, For the first Frequency domain representation of the scaled filter after frame update; For the first Frequency domain representation of the scaled filter after frame update; For the first The autocorrelation term of the frame model; For the first The autocorrelation term of the frame model; Preferably, step S3 specifically includes the following steps: S31. For each frame of image, the possible area of the beam edge line is offset to both sides by a certain range based on the line connecting the endpoints of the main beam after the update position. The specific definition of the offset to both sides of the line connecting the endpoints of the main beam is as follows: in, This is the pixel offset; pass , , , The enclosed polygonal area This represents the possible area along the edge of the beam. S32. Binarize the possible regions along the beam edge using the OStu method, with the following formula: in, No. The intensity value of the blue channel at the pixel location of a time-series panoramic image; The threshold that maximizes the inter-class variance of image gray levels; S33. The edge line of the main beam is extracted using the Canny algorithm, and then... The denoising criterion yields the main beam feature line, composed of a series of discrete pixel coordinates, as shown in the formula: in, For the set of edge lines; The coordinates of the edge line points; The coordinate point number; This represents the number of coordinate points on the edge line. Preferably, step S4 specifically includes the following steps: S41. Divide the main beam area into... The segments are continuous with second derivatives between each other; the fitted function is as follows: in, ; For, and satisfy C at all internal nodes 2 Continuity condition; for; for; for; S42. Calculate the single-pixel conversion coefficient using the following formula: in, Let be the length of the main beam, and let be the coordinates of the left end point of the main beam. The coordinates of the right endpoint are ; S43. Calculate the true alignment of the main beam using the following formula: in, For the first The actual alignment of the main beam in the frame; Before the bridge is in an unloaded state The average value of the frames; The main beam profile is represented in pixel coordinates; Preferably, step S5 specifically includes the following steps: S51, When the measured chord length is Midpoint The track live load chord measurement value at the point is given by the formula shown below: in, For the first frame dot The chord length is measured at the midpoint of the chord. For the first frame dot This is an approximation of the chord length measured at the midpoint of the chord. For the chord length, It is half the chord length; The k-th frame calculated in step S43 l The true line shape of a point; S52. The dynamic chord measurement value is obtained by superimposing the live load chord measurement value obtained in step S51 onto the static chord measurement value.
[0010] Therefore, the present invention employs the above-mentioned method for measuring the dynamic chord values of railway bridges based on UAV time-series images, which has the following beneficial effects: (1) This invention can measure the dynamic chord measurement value of a bridge during the passage of a train, solving the problem that traditional inertial navigation trolleys cannot measure the dynamic chord measurement value; (2) No operation or target installation is required on the bridge, and no drone is needed for panning and shooting, thus achieving true non-contact measurement; (3) This invention can be used as an important component of the dynamic chord measurement detection subsystem for railway bridges to evaluate the stiffness of railway bridges and provide support for railway bridge operation and maintenance decisions.
[0011] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0012] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a scene image of a drone hovering on the side of the main beam, as described in Embodiment 1 of the present invention. Figure 3The image shows a sequence of images of the main beam acquired by an unmanned aerial vehicle (UAV) in Embodiment 1 of the present invention. Figure 4 This is a view showing the selected end points of the main beam on both sides in Embodiment 1 of the present invention; Figure 5 This is the tracking position change curve of the two ends of the main beam in Embodiment 1 of the present invention; Figure 6 This is a binarized image of a sequence of images according to Embodiment 1 of the present invention; Figure 7 This is a diagram showing the characteristic edge line of the main beam in Embodiment 1 of the present invention; Figure 8 This is a fitting diagram of the main beam feature line in Embodiment 1 of the present invention; Figure 9 This is a comparison diagram of the measured deflection curve and the theoretical deflection curve of Embodiment 1 of the present invention; Figure 10 This is a comparison chart of the measured chord values and theoretical chord values of the main beam when the train is at mid-span, according to Embodiment 1 of the present invention. Figure 11 These are the measured chord values of the main beam of the train at different positions according to Embodiment 1 of the present invention; Figure 12 This is a scene image of a drone hovering on the side of the main beam, as shown in Embodiment 2 of the present invention. Figure 13 This refers to the sequence of images of the main beam acquired by an unmanned aerial vehicle (UAV) in Embodiment 2 of the present invention. Figure 14 This is a view showing the selected end points of the main beam on both sides in Embodiment 2 of the present invention; Figure 15 This is the tracking position change curve of the two ends of the main beam in Embodiment 2 of the present invention; Figure 16 This is a binarized image of the sequence image according to Embodiment 2 of the present invention; Figure 17 This refers to the edge feature line of the main beam in Embodiment 2 of the present invention; Figure 18 This is a fitting diagram of the main beam feature line in Embodiment 2 of the present invention; Figure 19 This is a comparison diagram of the measured deflection curve and the theoretical deflection curve in Embodiment 2 of the present invention; Figure 20 This is a comparison diagram of the measured chord values and theoretical chord values of the main beam when the train is at mid-span, according to Embodiment 2 of the present invention. Figure 21 These are the measured chord values of the main beam of the train at different positions in Embodiment 2 of the present invention. Detailed Implementation
[0013] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0014] This invention provides a method for measuring the dynamic ride comfort of railway bridges based on UAV time-series images. The method utilizes a UAV hovering over the mid-span side of the main girder to acquire time-series images of the main girder during train passage. Position and scale filters are used to track the beam end regions and are updated frame-by-frame. The updated beam end lines are offset to both sides by a certain range as potential edge regions of the main girder. OstU binarization and the Canny edge extraction algorithm are used to extract the bridge edges. The denoising criteria involve piecewise multi-form fitting to obtain a smooth curve. After subtracting the initial linear shape, a fixed chord length is used as the reference chord. The difference in moment between the midpoint and the reference chord is calculated point by point, and the dynamic ride comfort of the main beam during train passage is analyzed and calculated. Figure 1 As shown. The method specifically includes the following steps: S1. Acquire high-quality time-series images: Use a drone equipped with a stabilization gimbal to hover at a certain distance from the side of the main beam mid-span and collect time-series panoramic images of the entire process of the train from entering the bridge to completely leaving the bridge at a specific frame rate per second. The time-series panoramic images include the edge line of the main beam and the endpoints on both sides of the main beam. The time-series panoramic images include the edge line of the main beam and the endpoints on both sides of the main beam.
[0015] When recording images with a drone, an elevation angle of 1° to 3° is used, and the field of view includes clear beam feature lines or edge lines.
[0016] The formula for the acquired time-series panoramic images is: in, For indexing time-series panoramic images, ; These are the two-dimensional coordinates of pixels in the time-series panoramic image; For the first The RGB color vectors at pixel locations in a time-series panoramic image. ,in, This represents the intensity value of the red channel of the pixel. This represents the intensity value of the green channel of the pixel. This represents the intensity value of the blue channel for a given pixel.
[0017] S2. Locking and dynamically updating the beam end region: Using an interactive selection method, the beam end region of the main beam is selected with the two ends of the main beam as the center. The directional gradient histogram features and color features are extracted in the beam end region, and the position filter and scale filter are trained. For each subsequent frame, the position filter is used to detect around the previous frame to obtain a new position. The scale filter is used at the new position to detect the optimal scale. The position filter and scale filter are then updated. S21. Using an interactive selection method, select the beam end areas centered on both ends of the main beam, with the left end area of the main beam as... The right end area of the main beam is ,in, The x-coordinate of the top-left corner of the region. The ordinate of the top-left corner of the region. The width of the region, The height of the region; extract the directional gradient histogram features and color features of the main beam end region respectively, using the following formula: in, Features of the directional gradient histogram; Color characteristics; S22, Training Location Filter and Scale Filter: The objective function for training the location filter is set as minimizing a ridge regression problem, as shown in the formula: in, The weight vector to be learned; For in position The eigenvector at that location; For in position Expected output at; For regularization parameters, Weight vector The square of the L2 norm; The objective function is solved in the Fourier domain, as shown in the formula: in, For the Fourier transform of the filter; Fourier transform of the input features; The Fourier transform of the desired output; For regularization parameters; This indicates element-wise multiplication; , Indicates complex conjugation; Training the scaling filter: Dividing scale intervals: Defining the total number of scales as Using indexes The formula for distinguishing different scales is: ; The width and height of the target region are scaled according to a scale factor to obtain the region size for each scale, as shown in the formula: in, , These are the width and height of the target area, respectively. This is the scale factor, typically 1.02; This is a scale index, indicating the number of scaling points or levels relative to the base scale; Extracting scale features: For the target region at each scale, extract the feature vector, using the following formula: in, Scale features; , , These are the feature vectors at each scale; Define the training objective function: The objective function for training the scaling filter is set as minimizing the ridge regression problem, as shown in the formula: in, The weight coefficients to be learned The desired output is typically a Gaussian-shaped response. For regularization parameters; Weight vector The square of the L2 norm; Solving in the Fourier domain: Transform the objective function to the Fourier domain and solve it. The formula is as follows: in, This represents the scaling filter in the frequency domain. Fourier transform; For the expected output Fourier transform; scale features Fourier transform; This indicates element-wise multiplication; , To obtain the conjugate; For regularization parameters; S23. For each subsequent frame, a position filter is used to detect the area around the previous frame to obtain the new position; the position response is calculated according to the following formula: in, This is the location response result; The Fourier transform of the location feature vectors of the new candidate region; The autocorrelation term representing the model, For autocorrelation terms Fourier transform. Indicates the inverse Fourier transform; For the Fourier transform of the filter, take the complex conjugate; The scale response is calculated using the following formula: in, Fourier transform of the scale feature vectors of the new candidate region; The autocorrelation term represents the model; For the frequency domain representation of the scaling filter, take the complex conjugate; For the autocorrelation term of the model, take the complex conjugate; The updated version The left end region of the main beam of the frame is The right end area of the main beam is ; S24. At the new location, use a scaling filter to detect the optimal scale, and update the location filter and the scaling filter; the location filter update formula is: in, for The filter weights after frame update; for Frame filter weights; for The autocorrelation energy term after frame update; for The autocorrelation energy term of the frame; The learning rate; For time indexing; The scaling filter update formula is: in, for; for; for; for.
[0018] S3: Obtaining the clean edge line of the main beam: For each frame of the time-series panoramic image, based on the line connecting the endpoints of the main beam after the update position, a certain range is offset to both sides as the possible region of the beam's feature edge line. After binarizing the possible region of the beam's feature edge line using the OTU method, the Canny algorithm is used to extract the edge line of the main beam, and then... Criterion-based noise reduction; S31. For each frame of image, the possible area of the beam edge line is offset to both sides by a certain range based on the line connecting the endpoints of the main beam after the update position. The specific definition of the offset to both sides of the line connecting the endpoints of the main beam is as follows: in, This is the pixel offset; pass , , , The enclosed polygonal area This represents the possible area along the edge of the beam. S32. Binarize the possible regions along the beam edge using the OStu method, with the following formula: in, No. The intensity value of the blue channel at the pixel location of a time-series panoramic image; The threshold that maximizes the inter-class variance of image gray levels; S33. The edge line of the main beam is extracted using the Canny algorithm, and then... The denoising criterion yields the main beam feature line, composed of a series of discrete pixel coordinates, as shown in the formula: in, For the set of edge lines; The coordinates of the edge line points; The coordinate point number; This represents the number of coordinate points on the edge line.
[0019] S4. Obtain the true shape of the main beam: Fit the edge line of the main beam using a piecewise polynomial with continuous second derivative; calculate the image pixel conversion coefficient using a fixed beam length as a known dimension; multiply the piecewise polynomial fitting result with the pixel conversion coefficient to obtain the true shape of the main beam; S41. Divide the area within the main beam into... The segments are continuous with second derivatives between each other; the fitted function is as follows: in, ; For, and satisfy C at all internal nodes 2 Continuity condition; for; for; for; S42. Calculate the single-pixel conversion coefficient using the following formula: in, Let be the length of the main beam, and let be the coordinates of the left end point of the main beam. The coordinates of the right endpoint are ; S43. Calculate the true alignment of the main beam using the following formula: in, For the first The actual alignment of the main beam in the frame; Before the bridge is in an unloaded state The average value of the frames; The main beam profile is represented in pixel coordinates; S5. Obtain dynamic chord measurement values: Select the fixed chord length of the true alignment of the main beam as the reference chord. For each measuring point on the true alignment of the main beam, calculate the sagittal value of each measuring point relative to the reference chord to obtain the live load chord measurement value of the main beam. The live load chord measurement value of the main beam is superimposed with the static chord measurement value to obtain the dynamic chord measurement value. S51, When the measured chord length is Midpoint The track live load chord measurement value at the point is given by the formula shown below: in, For the first frame dot The chord length is measured at the midpoint of the chord. For the first frame dot This is an approximation of the chord length measured at the midpoint of the chord. For the chord length, It is half the chord length; The k-th frame calculated in step S43 l The true line shape of a point; S52. The dynamic chord measurement value is obtained by superimposing the live load chord measurement value obtained in step S51 onto the static chord measurement value.
[0020] Example 1 The implementation method of this invention is introduced by taking the dynamic chord measurement of a steel truss cable-stayed bridge as an example.
[0021] 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.
[0022] Step 1: Select a drone equipped with a stabilization gimbal and hover it at a certain distance from the side of the main beam mid-span. Capture time-series panoramic images of the train's entire journey from entering the bridge to completely leaving it at a specific frame rate per second. Figure 2 As shown. The time-series panoramic image should completely include the lower edge of the main beam and the endpoints on both sides of the main beam, as shown. Figure 3 As shown.
[0023] Step 2: Using an interactive selection method, select the beam end area of the main beam centered on the characteristic corner point, such as... Figure 4 As shown, histogram of oriented gradients and color features are extracted in the beam end region to train a position filter and a scale filter. For each subsequent frame, the position filter is used to detect the area around the previous frame to obtain a new position. At the new position, the scale filter is used to detect the optimal scale, updating the position and scale filters accordingly. Figure 5 As shown.
[0024] Step 3: For each frame of the image, offset a certain range to both sides of the line connecting the endpoints of the main beam after the update position as the possible region of the beam edge line. Binarize this possible region of the beam edge line using the OTU method, such as... Figure 6 As shown, the Canny algorithm is used to extract the edge line of the main beam, and then... Criterion-based noise reduction, such as Figure 7 As shown.
[0025] Step 4: Fit the edge line of the main beam using a piecewise polynomial with continuous second derivatives, such as... Figure 8 As shown, the fixed beam length is used as a known dimension to calculate the image pixel conversion coefficient, which is then multiplied by a piecewise polynomial scalar to obtain the true shape of the main beam.
[0026] Step 5: Using a fixed chord length as the reference chord for the actual alignment of the main beam, calculate the sagittal value for the reference chord length point by point, such as... Figure 9 As shown, further calculations of the dynamic chord measurement values of the main beam are performed, such as... Figure 10 and Figure 11 As shown.
[0027] Depend on Figure 10 and Figure 11 It can be seen that the maximum dynamic track chord measurement when the train passes occurs at the bridge tower, which is consistent with the theoretical analysis. The maximum chord measurement caused by the live load is 4.28 mm.
[0028] Example 2 The implementation of this invention is introduced by taking the dynamic chord measurement of a hybrid beam cable-stayed bridge as an example.
[0029] A certain cable-stayed bridge is a (67.5+60+60+350+60+60+67.5)m hybrid composite beam cable-stayed bridge with a total length of 726.8m. It adopts a hybrid main beam, with 209.4m of concrete box girders at both ends and a composite beam in the middle section. A steel-concrete composite section is set at the junction of the two. The main beam height is 4.5m.
[0030] Step 1: Select a drone equipped with a stabilization gimbal and hover it at a certain distance from the side of the main beam mid-span. Capture time-series panoramic images of the train's entire journey from entering the bridge to completely leaving it at a specific frame rate per second. Figure 12 As shown. The time-series panoramic image should completely include the lower edge of the main beam and the endpoints on both sides of the main beam, as shown. Figure 13 As shown.
[0031] Step 2: Using an interactive selection method, select the beam end area of the main beam centered on the characteristic corner point, such as... Figure 14 As shown, histogram of oriented gradients and color features are extracted in the beam end region to train a position filter and a scale filter. For each subsequent frame, the position filter is used to detect the area around the previous frame to obtain a new position. At the new position, the scale filter is used to detect the optimal scale, updating the position and scale filters accordingly. Figure 15 As shown.
[0032] Step 3: For each frame of the image, offset a certain range to both sides of the line connecting the endpoints of the main beam after the update position as the possible region of the beam edge line. Binarize this possible region of the beam edge line using the OTU method, such as... Figure 16 As shown, the Canny algorithm is used to extract the edge line of the main beam, and then... Criterion-based noise reduction, such as Figure 17 As shown.
[0033] Step 4: Fit the edge line of the main beam using a piecewise polynomial with continuous second derivatives, such as... Figure 18 As shown, the fixed beam length is used as a known dimension to calculate the image pixel conversion coefficient, which is then multiplied by a piecewise polynomial scalar to obtain the true shape of the main beam.
[0034] Step 5: Using a fixed chord length as the reference chord for the actual alignment of the main beam, calculate the sagittal value for the reference chord length point by point, such as... Figure 19 As shown, further calculations of the dynamic chord measurement values of the main beam are performed, such as... Figure 20 and Figure 21 As shown.
[0035] Depend on Figure 20 and Figure 21It can be seen that the maximum dynamic track chord measurement when the train passes occurs at the corresponding position of the bridge tower and mid-span, which is consistent with the theoretical analysis law. The maximum chord measurement caused by the live load is 6.62mm.
[0036] Therefore, this invention adopts the above-mentioned method for measuring the dynamic chord values of railway bridges based on UAV time-series images, using UAV non-contact acquisition as a foundation, and integrating position / scale filter tracking, OTU binarization, Canny edge extraction, and other methods. The core technologies, such as noise reduction and piecewise polynomial fitting, enable precise measurement of bridge dynamic chord values during train passage, filling the gap in dynamic acceptance testing. Furthermore, they eliminate the need for any sensors or targets on the bridge, making them suitable for long-span railway bridges in complex environments such as river crossings, mountain valleys, and other challenging terrains, significantly reducing measurement difficulty and cost. Simultaneously, their high measurement frequency and strong data continuity provide accurate and reliable deformation data support for railway bridge stiffness assessment, dynamic ride comfort testing, and maintenance decisions, offering a novel technological approach for the field of dynamic railway bridge testing.
[0037] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for measuring the dynamic chord values of railway bridges based on UAV time-series images, characterized in that, Includes the following steps: S1. Acquire high-quality time-series images: Use a drone equipped with a stabilization gimbal to hover on the side of the main beam mid-span and collect time-series panoramic images of the train from entering the bridge to completely leaving the bridge at a specific frame rate per second. S2. Locking and dynamically updating the beam end region: Using an interactive selection method, the beam end region of the main beam is selected with the two ends of the main beam as the center. The directional gradient histogram features and color features are extracted in the beam end region, and the position filter and scale filter are trained. For each subsequent frame, the position filter is used to detect around the previous frame to obtain a new position. The scale filter is used at the new position to detect the optimal scale. The position filter and scale filter are then updated. S3: Obtaining the clean edge line of the main beam: For each frame of the time-series panoramic image, based on the line connecting the endpoints of the main beam after the update position, a certain range is offset to both sides as the possible region of the beam's feature edge line. After binarizing the possible region of the beam's feature edge line using the OTU method, the Canny algorithm is used to extract the edge line of the main beam, and then... Criterion-based noise reduction; S4. Obtain the true shape of the main beam: Fit the edge line of the main beam using a piecewise polynomial with continuous second derivative; calculate the image pixel conversion coefficient using a fixed beam length as a known dimension; multiply the piecewise polynomial fitting result with the pixel conversion coefficient to obtain the true shape of the main beam; S5. Obtain dynamic chord measurement values: Select the fixed chord length of the true alignment of the main beam as the reference chord. For each measuring point on the true alignment of the main beam, calculate the sagittal value of each measuring point relative to the reference chord to obtain the live load chord measurement value of the main beam. The live load chord measurement value of the main beam is superimposed with the static chord measurement value to obtain the dynamic chord measurement value.
2. The method for measuring dynamic chord values of railway bridges based on UAV time-series images according to claim 1, characterized in that, The time-series panoramic image in step S1 includes the edge line of the main beam and the endpoints on both sides of the main beam.
3. The method for measuring dynamic chord values of railway bridges based on UAV time-series images according to claim 1, characterized in that, In step S1, when the drone records images, it uses an elevation angle of 1° to 3°, and the field of view includes the feature lines or edge lines of the beam.
4. The method for measuring dynamic chord values of railway bridges based on UAV time-series images according to claim 2, characterized in that, The time-series panoramic images acquired in step S1 are represented by the following formula: in, For indexing time-series panoramic images, ; These are the two-dimensional coordinates of pixels in the time-series panoramic image; For the first The RGB color vectors at pixel locations in a time-series panoramic image. ,in, This represents the intensity value of the red channel of the pixel. This represents the intensity value of the green channel of the pixel. This represents the intensity value of the blue channel for a given pixel.
5. The method for measuring dynamic chord values of railway bridges based on UAV time-series images according to claim 4, characterized in that, Step S2 specifically includes the following steps: S21. Using an interactive selection method, select the beam end areas centered on both ends of the main beam, with the left end area of the main beam as... The right end area of the main beam is ,in, The x-coordinate of the top-left corner of the region. The ordinate of the top-left corner of the region. The width of the region, The height of the region; extract the directional gradient histogram features and color features of the main beam end region respectively, using the following formula: in, Features of the directional gradient histogram; Color characteristics; S22, Training Location Filter and Scale Filter: The objective function for training the location filter is set as minimizing a ridge regression problem, as shown in the formula: in, The weight vector to be learned; For in position The eigenvector at that location; For in position Expected output at; For regularization parameters, Weight vector The square of the L2 norm; The objective function is solved in the Fourier domain, as shown in the formula: in, For the Fourier transform of the filter; Fourier transform of the input features; The Fourier transform of the desired output; For regularization parameters; This indicates element-wise multiplication; , Indicates complex conjugation; Training the scaling filter: Dividing scale intervals: Defining the total number of scales as Using indexes The formula for distinguishing different scales is: ； The width and height of the target region are scaled according to a scale factor to obtain the region size for each scale, as shown in the formula: in, , These are the width and height of the target area, respectively. Scale factor; This is a scale index, indicating the number of scaling points or levels relative to the base scale; Extracting scale features: For the target region at each scale, extract the feature vector, using the following formula: in, Scale features; , , These are the feature vectors at each scale; Define the training objective function: The objective function for training the scaling filter is set as minimizing the ridge regression problem, as shown in the formula: in, The weight coefficients to be learned For the desired output, For regularization parameters; Weight vector The square of the L2 norm; Solving in the Fourier domain: Transform the objective function to the Fourier domain and solve it. The formula is as follows: in, This represents the scaling filter in the frequency domain. Fourier transform; For the expected output Fourier transform; scale features Fourier transform; This indicates element-wise multiplication; , To obtain the conjugate; For regularization parameters; S23. For each subsequent frame, a position filter is used to detect the area around the previous frame to obtain the new position; the position response is calculated according to the following formula: in, This is the location response result; The Fourier transform of the location feature vectors of the new candidate region; The autocorrelation term representing the model, For autocorrelation terms Fourier transform; Indicates the inverse Fourier transform; For the Fourier transform of the filter, take the complex conjugate; The scale response is calculated using the following formula: in, Fourier transform of the scale feature vectors of the new candidate region; The autocorrelation term represents the model; For the frequency domain representation of the scaling filter, take the complex conjugate; For the autocorrelation term of the model, take the complex conjugate; The updated version The left end region of the main beam of the frame is The right end area of the main beam is ; S24. At the new location, use a scaling filter to detect the optimal scale, and update the location filter and the scaling filter; the location filter update formula is: in, for The filter weights after frame update; for Frame filter weights; for The autocorrelation energy term after frame update; for The autocorrelation energy term of the frame; The learning rate; For time indexing; The scaling filter update formula is: in, For the first Frequency domain representation of the scaled filter after frame update; For the first Frequency domain representation of the scaled filter after frame update; For the first The autocorrelation term of the frame model; For the first The autocorrelation term of the frame model.
6. The method for measuring dynamic chord values of railway bridges based on UAV time-series images according to claim 5, characterized in that, Step S3 specifically includes the following steps: S31. For each frame of image, the possible area of the beam edge line is offset to both sides by a certain range based on the line connecting the endpoints of the main beam after the update position. The specific definition of the offset to both sides of the line connecting the endpoints of the main beam is as follows: in, This is the pixel offset; pass , , , The enclosed polygonal area This represents the possible area along the edge of the beam. S32. Binarize the possible regions along the beam edge using the OStu method, with the following formula: in, No. The intensity value of the blue channel at the pixel location of a time-series panoramic image; The threshold that maximizes the inter-class variance of image gray levels; S33. The edge line of the main beam is extracted using the Canny algorithm, and then... The denoising criterion yields the main beam feature line, composed of a series of discrete pixel coordinates, as shown in the formula: in, For the set of edge lines; The coordinates of the edge line points; The coordinate point number; This represents the number of coordinate points on the edge line.
7. The method for measuring dynamic chord values of railway bridges based on UAV time-series images according to claim 6, characterized in that, Step S4 specifically includes the following steps: S41. Divide the area within the main beam into... The segments are continuous with second derivatives between each other; the fitted function is as follows: in, ; Let C be the fitted function, and satisfy C at all internal nodes. 2 Continuity condition; It is a piecewise function; The coefficients are those of the piecewise function. for of Power; S42. Calculate the single-pixel conversion coefficient using the following formula: in, Let be the length of the main beam, and let be the coordinates of the left end point of the main beam. The coordinates of the right endpoint are ; S43. Calculate the true alignment of the main beam using the following formula: in, For the first The actual alignment of the main beam in the frame; Before the bridge is in an unloaded state The average value of the frames; The main beam profile is represented in pixel coordinates.
8. The method for measuring dynamic chord values of railway bridges based on UAV time-series images according to claim 7, characterized in that, Step S5 specifically includes the following steps: S51, When the measured chord length is Midpoint The track live load chord measurement value at the point is given by the formula shown below: in, For the first frame dot The chord length is measured at the midpoint of the chord. For the first frame dot This is an approximation of the chord length measured at the midpoint of the chord. For the chord length, It is half the chord length; The k-th frame calculated in step S43 l The true line shape of a point; S52. The dynamic chord measurement value is obtained by superimposing the live load chord measurement value obtained in step S51 onto the static chord measurement value.