A non-target cable force identification method based on edge recognition

Abstract

The application discloses a kind of based on edge identification no target cable force identification method, the method includes: cable video acquisition and pre-processing: acquisition cable video, video frame is carried out image gray conversion and is adapted to the multi-scale Gaussian filtering operation of filter scale according to noise characteristics;Canny algorithm cable edge identification: using Canny edge detection algorithm to the video frame after pre-processing is carried out cable edge identification;Cable feature point screening and KLT optical flow method identification: from edge identification result screening cable feature point, and using KLT optical flow method in continuous video frame between tracking feature point dynamic displacement;Fundamental frequency identification and cable force calculation: displacement data is applied Fourier transform to identify cable fundamental frequency, and adopts XGBOOST regression model based on fundamental frequency and relevant parameter calculation cable tension.The application realizes the precise, efficient and non-contact monitoring of bridge cable stress state.

Description

Technical Field

[0001] This invention belongs to the field of image recognition and intelligent sensing technology, and in particular relates to a targetless cable force recognition method based on edge recognition. Background Technology

[0002] In bridge engineering, cable tension is a key parameter for assessing structural health. Traditional cable tension detection methods mainly rely on the installation of contact sensors and manual data collection, which is not only cumbersome and inefficient, but also susceptible to environmental interference, leading to unstable data accuracy. With the development of infrastructure construction in my country, the requirements for the accuracy and efficiency of bridge health monitoring are increasing, and this detection mode relying on manual labor and contact sensors can no longer meet the needs of modern bridge engineering.

[0003] In recent years, the rapid development of digital information technologies such as digital twins and computer vision has provided new technological pathways for non-contact cable force detection. Several industry plans issued by the Ministry of Transport have also explicitly proposed promoting the deep integration of next-generation information technology with bridge engineering to improve the intelligence level of structural monitoring. However, existing non-contact cable force identification technologies still have significant limitations: on the one hand, existing methods are highly sensitive to environmental changes and easily affected by external factors such as lighting and weather, leading to unstable cable force identification accuracy; on the other hand, existing methods suffer from a contradiction between real-time performance and computational resource requirements, with complex calculation processes causing system response delays and making it difficult to meet the real-time requirements of dynamic monitoring.

[0004] Furthermore, most existing non-contact cable force identification methods require manual placement of clearly visible reflective targets on the cables, which increases the complexity and safety risks of on-site operations. Target-free cable force identification technology still faces significant challenges in terms of accuracy and robustness, especially in complex bridge environments where it is difficult to accurately extract cable motion information from the background. Simultaneously, cable force calculation methods based on traditional mathematical formulas often rely on simplified physical models, neglecting actual factors such as cable elastic deformation and boundary constraint changes, leading to significant deviations between calculated results and actual cable forces.

[0005] Therefore, there is a need to develop a non-contact cable force identification method that can adapt to complex environments, achieve real-time and accurate identification, eliminate the need for manual targets, and overcome the limitations of traditional calculation models, so as to meet the requirements of modern bridge health monitoring for accuracy, efficiency and reliability. Summary of the Invention

[0006] The purpose of this invention is to provide a targetless cable force identification method based on edge recognition, which solves the problems of insufficient cable force identification accuracy and poor real-time performance caused by environmental interference sensitivity, low computational efficiency, reliance on manual targets, and oversimplification of traditional physical models in the prior art, and realizes accurate, efficient and non-contact monitoring of the stress state of bridge cables.

[0007] The technical solution adopted in this invention is a targetless cable force recognition method based on edge recognition, comprising the following steps:

[0008] Step S1: Acquire cable video, perform image grayscale conversion operation on video frames in sequence, and perform multi-scale Gaussian filtering operation that adapts the filtering scale according to noise characteristics;

[0009] Step S2: Use the Canny edge detection algorithm to perform cable edge recognition on the preprocessed video frames; the Canny edge detection algorithm includes image smoothing, gradient calculation, non-maximum suppression, double thresholding, and edge connection;

[0010] Step S3: Filter cable feature points from the Canny edge recognition results, including identifying long continuous edge segments and extracting their midpoints as displacement tracking markers, and supplementing by extracting significant corner points and curvature change points on the edges; use the KLT optical flow method to track the dynamic displacement of the feature points between consecutive video frames;

[0011] Step S4: Apply Fourier transform to the displacement data obtained by displacement tracking to convert it into a frequency domain signal to identify the fundamental frequency of the cable, and use a machine learning model to calculate the cable tension based on the fundamental frequency and related parameters.

[0012] Furthermore, the image smoothing in S2 employs Gaussian filtering, and the gradient calculation uses the Sobel operator to calculate the horizontal and vertical gradient components of the image.

[0013] Furthermore, the dual thresholding in S2 employs an adaptive thresholding method, including calculating the mean gradient magnitude within a local region. and standard deviation And dynamically calculate the adaptive high threshold according to the formula. Low threshold :

[0014] ;

[0015] ;

[0016] in, To adjust parameters; mean Indicates the overall gradient level of the region; standard deviation This indicates gradient volatility.

[0017] Furthermore, the adjustment parameters The optimal value is found within a preset range [0.5, 2.5] using a binary search method.

[0018] Furthermore, the non-maximum suppression in S2 is achieved as follows: after calculating the gradient magnitude and direction, the gradient magnitude of each pixel is compared with the gradient magnitude of its two adjacent pixels along the gradient direction; if it is not a local maximum, the gradient magnitude of that pixel is set to zero.

[0019] The edge connection is achieved by taking the strong edge pixel obtained after double thresholding as the starting point and checking the weak edge pixels in its eight neighborhoods; if the weak edge pixel is connected to any strong edge pixel, it is promoted to a strong edge, otherwise it is set to zero, so as to obtain a continuous edge image.

[0020] Furthermore, the KLT optical flow method in S3 is based on the assumptions of constant brightness, small motion displacement, and spatial consistency.

[0021] Furthermore, based on the aforementioned spatial consistency assumption, the displacement vector of the feature point is calculated by solving the optical flow equation using the least squares method. ;

[0022] in, , These are optical flow vectors, collectively representing the motion vectors of the cable edge feature points between adjacent frames. This represents the horizontal component of the motion vector. This represents the vertical component of the motion vector.

[0023] Furthermore, the KLT optical flow method in S3 employs an image pyramid structure, in which the pyramid progressively reduces the resolution by decreasing the sampling factor from the bottom layer (original image) to the top layer, in order to capture large displacements at the coarse resolution layer and refine them to the fine resolution layer.

[0024] Furthermore, the fundamental frequency identification in S4 includes identifying the dominant vibration frequency with the largest amplitude in the frequency domain signal as the fundamental frequency.

[0025] Furthermore, the machine learning model in S4 is an XGBOOST regression model, and the relevant parameters include cable length and line density.

[0026] The beneficial effects of this invention are:

[0027] 1. This invention significantly improves the accuracy of cable image feature extraction and displacement tracking by introducing the Canny edge detection algorithm. This algorithm can accurately extract clear cable contour edges, providing stable and high-quality feature point input for subsequent KLT optical flow methods. It effectively overcomes displacement tracking errors caused by blurred edges or inaccurate feature point extraction in traditional methods, achieving accurate identification of cable vibration displacement.

[0028] 2. This invention enhances the applicability and robustness in complex environments by integrating an automatic threshold adjustment mechanism. This mechanism can dynamically optimize the detection threshold of the Canny algorithm based on conditions such as ambient light and noise levels, achieving adaptive extraction of cable edges. This avoids the problem of unstable recognition results under different working conditions caused by the fixed threshold method, thus ensuring the reliability of the cable force recognition system in variable field environments.

[0029] 3. This invention uses the XGBOOST regression prediction model to correct the cable force calculation results, which significantly improves the accuracy of cable force calculation. At the same time, the combination of Canny edge detection and improved KLT optical flow method optimizes the calculation process, which improves the overall computing efficiency while ensuring recognition accuracy and reduces system resource dependence. This makes the method applicable to real-time monitoring scenarios of various bridge cables and broadens the application scope of targetless cable force recognition technology. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the method of the present invention.

[0032] Figure 2 This is a schematic diagram of Canny edge adaptive thresholding.

[0033] Figure 3 This is a diagram comparing various edge detection algorithms.

[0034] Figure 4 This is a schematic diagram of the optical flow image pyramid principle.

[0035] Figure 5 This is a schematic diagram of an experimental setup for measuring cable force based on the frequency method.

[0036] Figure 6The graphs show the performance verification results of the cable force prediction model based on the XGBOOST algorithm. (a) is a scatter plot of the predicted and true values ​​in the training set, (b) is a scatter plot of the predicted and true values ​​in the test set, (c) is a curve comparison of the predicted and true values ​​in the training set, and (d) is a curve comparison of the predicted and true values ​​in the test set.

[0037] Figure 7 This is a schematic diagram of the setup for a real bridge test to detect cable tension.

[0038] Figure 8 These are comparison charts of cable vibration identification results, where (a) is the displacement time history curve identified by the present invention, and (b) is the acceleration time history curve measured by the sensor.

[0039] Figure 9 The image shows a comparison of the cable vibration spectrum identification results, where (a) is the spectrum identification result of the present invention and (b) is the actual measured spectrum data of the sensor. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] like Figure 1 As shown in the embodiment, a targetless cable force recognition method based on edge recognition is provided, which includes the following steps:

[0042] Step S1: Cable video acquisition and preprocessing: High-definition camera equipment is used to acquire cable video; image grayscale conversion is performed on the acquired video frames to convert color video frames into grayscale frames; multi-scale Gaussian filtering algorithm is used to filter the grayscale converted video frames, and the filtering scale is adapted according to different noise characteristics to suppress random noise in the video while preserving key details such as cable edges, thus completing the preprocessing of the cable video.

[0043] Step S2: Canny algorithm for cable edge recognition:

[0044] The Canny edge detection algorithm is an edge detection method used to identify cable edges. The algorithm's main process includes five steps: image smoothing, gradient calculation, non-maximum suppression, double thresholding, and edge connection. It aims to improve the accuracy of edge detection and reduce the generation of false edges.

[0045] S21: Canny Edge Detection Process:

[0046] Gaussian filtering is used to smooth images and reduce noise. The Gaussian filter smooths the image through a convolution operation, and its formula is as follows:

[0047]

[0048] in, Represents the Gaussian function; y is the standard deviation of the Gaussian distribution, which controls the smoothness of the filter; x and y are the horizontal and vertical coordinate offsets of any point in the Gaussian convolution kernel relative to its center point. For example, for a 3×3 convolution kernel, the values ​​of x and y are both in the range of {-1, 0, 1}.

[0049] The gradient of the image is calculated to find regions of abrupt brightness changes (i.e., edges). The Sobel operator is used to calculate the horizontal gradient component of the image. and vertical gradient components The magnitude of the gradient and direction Calculate using the following formulas respectively:

[0050]

[0051] The gradient magnitude reflects the magnitude of pixel value changes, while the gradient direction indicates the direction of the edge. Edges in an image typically appear where the gradient value is large, i.e., where the image brightness changes most significantly.

[0052] After obtaining the gradient information, non-maximum suppression is introduced to refine the edges and remove artifacts. This step compares the gradient magnitude of each pixel with the gradient magnitude of its two adjacent pixels along the gradient direction, and sets the value to zero if it is not a local maximum, thus obtaining a finer edge.

[0053] Double thresholding and edge connectivity are performed. Pixels are divided into three categories by setting two thresholds: high and low thresholds. Pixels with gradient values ​​higher than the high threshold are marked as strong edges, those between the high and low thresholds are marked as weak edges, and those below the low threshold are considered non-edges. Strong edge pixels are set to their maximum value, while non-edge pixels are set to zero.

[0054] Edge connectivity employs a "hysteresis method" to ensure continuity: starting with a strong edge pixel, weak edge pixels within its eight-neighborhood are examined; if a weak edge pixel is connected to any strong edge, it is promoted to a strong edge; otherwise, it is set to zero. This process eliminates isolated weak edges through connectivity analysis, ultimately yielding a continuous and complete edge image. Traditional algorithms use fixed high and low thresholds, but this invention employs an adaptive method, as described in S22, to improve performance in complex images.

[0055] S22: Improved Adaptive Threshold:

[0056] To overcome the limitations of the traditional Canny algorithm with a fixed threshold in complex images (such as scenes with uneven brightness or high noise), this invention employs an adaptive threshold-improved Canny edge detection algorithm. Figure 2 As shown, a sliding window (such as 7×7) is used to perform local region calculations on the image to obtain gradient information for each region.

[0057] Calculate the mean of the gradient magnitude G within the local region. and standard deviation The formula is as follows:

[0058]

[0059]

[0060] Among them, the mean Indicates the overall gradient level of the region; standard deviation This indicates gradient volatility; the larger the standard deviation, the more drastic the edge changes. N is the total number of pixels within the window. This represents the gradient magnitude.

[0061] An adaptive high threshold is dynamically calculated based on the gradient mean and standard deviation of the local region. Low threshold :

[0062]

[0063]

[0064] The parameter k is used to control the sensitivity of the threshold: increasing the k value reduces the number of detected strong edges but increases reliability; decreasing the k value retains more weak edges but increases the possibility of false noise. Based on the calculated threshold, pixels with gradient magnitudes greater than or equal to the high threshold are classified as strong edges; pixels between the low and high thresholds are marked as weak edges; and pixels below the low threshold are considered non-edges and suppressed. To further optimize performance, this invention introduces a parameter search method based on edge connectivity: edge detection is performed using different k values, and the edge continuity is evaluated (quantified by indicators such as the average length of edge segments); the optimal k value is found within a preset range [0.5, 2.5] using a binary search method. This optimal solution aims to minimize detected edge breaks and achieve appropriate noise suppression, thereby realizing adaptive threshold selection and improved detection accuracy.

[0065] S23: Algorithm Performance Evaluation

[0066] To evaluate the accuracy and applicability of the proposed adaptive threshold-improved Canny edge detection algorithm in tethered body images, this study selected multiple sets of photos containing different background conditions and various types of tethered bodies as test samples. By extracting edges from the sample images and comparing them with manual annotations, the detection performance of several commonly used algorithms, including the proposed Canny, Laplacian, LSD, Prewitt, and Sobel algorithms, under different lighting conditions, texture complexity, and tethered body morphologies was systematically compared.

[0067] like Figure 3 As shown, the compared algorithms each have their own characteristics: the Laplacian operator, based on the second derivative, can highlight areas with rapid changes in image grayscale, but is extremely sensitive to noise; the LSD algorithm focuses on line segment extraction and is suitable for structured geometric feature recognition in complex backgrounds; the Prewitt operator achieves edge detection through gradient calculation in the horizontal and vertical directions, is simple and computationally efficient, but has weak noise resistance; while the Sobel operator gives greater weight to the center pixel in gradient calculation, which can more effectively suppress noise and extract details. By comparing the performance of each algorithm in edge sharpness, continuity, and noise suppression, a scientific basis and reference are provided for the selection of edge detection algorithms in actual cable monitoring.

[0068] To evaluate the performance of the improved Canny algorithm in cable edge recognition, Precision, Recall, F1-score, and IoU were selected as evaluation metrics, and compared with algorithms such as Laplacian, LSD, Prewitt, and Sobel, as shown in Table 1.

[0069] Precision refers to the proportion of samples that are actually positive out of all samples predicted to be positive. The formula for calculation is:

[0070]

[0071] TP represents the true positives, which are correctly predicted as positive samples; FP represents the false positives, which are actually negative samples but were incorrectly predicted as positive samples.

[0072] Recall refers to the proportion of all actual positive samples that are correctly predicted as positive samples. The formula is:

[0073]

[0074] FN represents the number of false negatives, which are incorrectly predicted as negative samples.

[0075] The F1 score is the harmonic mean of precision and recall, used to balance the two. The formula is:

[0076]

[0077] IoU represents the degree of overlap between the predicted region and the actual region, and is calculated using the following formula:

[0078]

[0079] in, To predict the area of ​​intersection between the actual region and the real region, The area is the union of the predicted region and the actual region.

[0080] Table 1 Comparison of Edge Recognition Accuracy of Edge Recognition Algorithms

[0081]

[0082] Evaluation results show that the improved Canny algorithm has significant advantages in cable edge recognition: its precision reaches 0.8572, far higher than other algorithms, with fewer false positives; its recall is 0.9905, at a high level, with fewer missed detections; its F1-score of 0.9015 far exceeds other algorithms, achieving a better balance between accuracy and completeness; and its IoU of 0.7566 is also higher than other algorithms, indicating better edge matching. In summary, compared with algorithms such as Laplacian, LSD, Prewitt, and Sobel, the improved Canny algorithm has significant advantages in accuracy, overall performance, and edge matching in cable edge recognition, making it more suitable for edge feature extraction in targetless cable force recognition.

[0083] S24: Cable Feature Point Selection: For the cable edge image obtained through Canny edge recognition, representative feature points are selected based on pixel gradients and geometric features to serve as markers for subsequent displacement tracking. Long, continuous edge segments in the image (e.g., longer than 100 pixels) are identified, and their midpoints are extracted as primary displacement tracking markers. This selection is based on the highest displacement tracking stability of the midpoints. To enhance the discriminative power of the feature point set, significant corner points and curvature abrupt change points on each edge are extracted. Finally, a feature point set with both high stability and high discriminative power is output.

[0084] Step S3: KLT optical flow identification: KLT optical flow is used to track the dynamic displacement of cable edge feature points between consecutive video frames.

[0085] S31: Algorithm Principles and Assumptions:

[0086] This optical flow method is based on three core assumptions:

[0087] (1) Brightness constant assumption: The brightness of the same feature point remains unchanged between adjacent frames, that is:

[0088]

[0089] in, This represents the brightness of pixel (x, y) at time t. and For displacement, This represents the time interval between adjacent video frames.

[0090] (2) Small motion displacement assumption: The displacement of feature points between adjacent frames is very small. Based on this, a first-order Taylor expansion is performed on the right side of the constant brightness assumption equation, and higher-order infinitesimal terms are ignored, resulting in:

[0091]

[0092] Based on the assumption of constant brightness, the fundamental equation of optical flow can be derived:

[0093]

[0094] in, , , The images are respectively in , Gradient in direction and time-domain derivative; , These are optical flow vectors, collectively representing the motion vectors of the cable edge feature points between adjacent frames. This represents the horizontal component of the motion vector. This represents the vertical component of the motion vector.

[0095] (3) Spatial consistency assumption: Adjacent pixels within a local window have similar motion patterns. Based on this, the displacement vectors of feature points are calculated by solving the overdetermined equations using the least squares method. :

[0096]

[0097] By minimizing error We obtain a closed-form solution:

[0098]

[0099] in, , The gradient matrix A consists of the image spatial gradients (i.e., the rate of change of brightness in space) of the i-th pixel in the x and y directions. The temporal derivative of the image at the i-th pixel (i.e., the brightness change of that pixel between adjacent frames) forms the temporal derivative vector b. This represents the total number of pixels within the local window used when calculating the displacement of feature points; A is the gradient matrix; and b is the time-domain derivative vector.

[0100] S32: Image Pyramid and Implementation Process: To effectively handle motion at different scales, the algorithm adopts an image pyramid structure, such as... Figure 4 As shown, the pyramid typically downsamples from the bottom (Level 0, the original image) to the top (Level 3) by a factor of 1 / 2, gradually reducing the resolution layer by layer. This helps capture large displacements at the coarse-resolution layer, and then refines the image to a fine-resolution layer through upsampling and iteration, thus balancing computational efficiency and tracking accuracy. Its core process involves matching selected feature points across consecutive video frames, calculating their displacement vectors in adjacent frames, and then recording the displacement coordinates and motion trajectory data of each feature point.

[0101] Step S4: Fundamental frequency identification and cable force calculation: This step converts the displacement data obtained by the optical flow method into a frequency domain signal to identify the fundamental frequency, and uses a machine learning model to calculate the cable force.

[0102] S41: Baseband Identification:

[0103] Fourier transform was applied to the cable displacement trajectory data (time domain signal) acquired by S3 to convert it into a frequency domain signal. By analyzing the spectrum, different frequency components were decomposed, and the principal vibration frequency with the most significant energy was determined. The principal vibration frequency with the largest amplitude in the frequency domain signal was identified as the fundamental frequency and used as the fundamental frequency of the cable.

[0104] S42: Cable Force Calculation and XGBOOST Model Correction:

[0105] To overcome the computational bias introduced by the simplification assumptions in traditional frequency-based physical formulas, this step introduces the XGBOOST regression prediction model to correct the cable force calculation method. This model, through learning from a large amount of experimental data, constructs a nonlinear mapping relationship between cable parameters (fundamental frequency, cable length, linear density) and actual cable force, without relying on a pre-defined simplified physical model. By substituting the fundamental frequency and related parameters identified by S41 into this trained model, high-precision cable tension values ​​can be directly calculated and output, effectively improving the accuracy and reliability of calculations under complex boundary conditions.

[0106] The core advantage of the XGBOOST algorithm lies in its efficient and precise loss function optimization mechanism. Building upon the traditional Gradient Boosting Tree (GBDT) approach, it performs a second-order Taylor expansion of the loss function and utilizes first-order gradient and second-order Hessian matrix information to optimize model parameters. This allows for more precise loss reduction at each iteration, significantly improving the model's convergence speed and prediction accuracy. This optimization mechanism, combined with a built-in regularization strategy, introduces a penalty term for model complexity (such as the number of leaf nodes and node weights) into the objective function, effectively suppressing overfitting. This enables the model to possess both high fitting ability and strong generalization performance when dealing with complex nonlinear relationships between force and parameters.

[0107] like Figure 5 As shown, to train the model, three steel cables with lengths of 5.5m, 6.8m, and 8.3m were used as test objects in the laboratory. A hydraulic tensioner was used to conduct 240 tests within the range of 50N to 4000N. Cable length, mass per unit length, and fundamental frequency were used as model input features, and cable force was used as the model prediction target. The 240 sets of test data were divided in a 7:3 ratio, with 168 sets (70%) used as the training set and 72 sets (30%) used as the test set.

[0108] Table 2 Comparison of predicted cable force values ​​for the test set in the experimental section

[0109]

[0110] As shown in Table 2 and Figure 6 As shown in (a)-(d), the trained model exhibits excellent predictive performance: in terms of goodness of fit, the coefficient of determination R² for the training set is as high as 0.999, and the R² for the test set is 0.996, indicating that the model fits the data patterns very well and has stable generalization ability. Regarding prediction error, the prediction error of cable force under different cable lengths is controlled within 1.3% (0.92% error for 5.5m cable, 1.07% error for 6.8m cable, and 1.24% error for 8.3m cable).

[0111] The above results fully verify that the XGBOOST regression model can accurately capture the complex nonlinear relationship between cable physical parameters and cable force, and can be used as a reliable method for efficient and accurate prediction of cable force.

[0112] To verify the effectiveness and accuracy of the targetless cable force identification method proposed in this invention, a comparative experiment was conducted in a real bridge scenario.

[0113] Test setup as follows Figure 7As shown, three cables, C5, C7, and C8, of the bridge were selected as test objects. Rubber hammers were used to artificially excite the cables to generate vibration. For comparison, an accelerometer connected to a DASP-v11 dynamic data acquisition instrument from Dongfang Institute was used as the benchmark, positioned in the lower middle of the cable, with a sampling frequency of 200Hz. The image acquisition camera of the method described in this invention was set up 6 meters away from the cable, with a sampling frequency of 50Hz.

[0114] To verify the accuracy of the method of the present invention, the displacement identification results based on optical flow method were simultaneously compared with the measurement data of the accelerometer. Figure 8 (a) and Figure 8 (b) The cable displacement time history curve identified by the method of the present invention and the acceleration time history curve collected by the acceleration sensor are respectively. Spectral analysis is performed on the above time history data to obtain their respective spectra, i.e. Figure 9 (a) and Figure 9 (b) The results show that the method of the present invention identifies the first-order frequency of the cable as 5.128 Hz and the second-order frequency as 10.307 Hz; the accelerometer identifies the first-order frequency as 5.204 Hz and the second-order frequency as 10.308 Hz. The results are highly similar, with the calculation errors for both the first and second-order frequencies being less than 0.2%, fully verifying the accuracy and reliability of the method of the present invention in frequency identification.

[0115] Table 3. Experimental Results

[0116]

[0117] Table 3 lists the specific frequency values ​​and error analysis results. For cables C5, C7, and C8, the maximum error between the first-order frequency identified by the method of this invention and the accelerometer results is 0.33%, and the maximum error for the second-order frequency is 0.48%. All frequency identification errors are controlled within 0.5%.

[0118] This invention aims to overcome the limitations of traditional measurement techniques in complex environments, achieving high-precision non-contact detection of bridge cable forces and providing key technical support for real-time bridge health monitoring. Addressing the problem of large displacement tracking errors caused by blurred cable edges and inaccurate feature point extraction in targetless cable force identification using the traditional KLT optical flow method, this invention innovatively introduces the Canny edge detection method in the image preprocessing stage. By accurately extracting clear and continuous edge features of the cables, it provides stable and reliable feature point input for optical flow tracking, thereby significantly improving the accuracy and continuity of displacement tracking. Furthermore, to solve the problem of unstable edge detection caused by a fixed threshold in the traditional Canny algorithm under complex lighting and noise environments, this invention proposes an adaptive threshold adjustment mechanism based on image histogram distribution and gradient magnitude characteristics. This mechanism can dynamically optimize the threshold setting according to the actual image content, ensuring reliable extraction of cable edges under different environmental conditions and enhancing the robustness and applicability of the method. Furthermore, addressing the shortcomings of traditional frequency-based cable force calculations due to errors introduced by simplified model assumptions, this invention introduces the XGBOOST regression model. Leveraging its powerful nonlinear learning capabilities, it constructs a mapping relationship between cable parameters and actual cable forces, enabling data-driven correction of traditional calculation results and effectively improving cable force identification accuracy. Through the organic integration of these multiple improvements, this invention effectively enhances the accuracy, environmental adaptability, and engineering practical value of targetless cable force identification, demonstrating promising application prospects.

[0119] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A non-target-cable force identification method based on edge identification, characterized in that, The method comprises the following steps: Step S1: collecting a cable video, performing image grayscale conversion operation on video frames in sequence, and performing multi-scale Gaussian filtering operation according to noise characteristics to adapt filtering scale; Step S2: performing cable edge identification on the preprocessed video frames using a Canny edge detection algorithm; the Canny edge detection algorithm comprises image smoothing, gradient calculation, non-maximum suppression, double thresholding and edge connection; Step S3: screening cable feature points from the Canny edge identification result, comprising determining long continuous edge segments and extracting points therein as displacement tracking marker points, and additionally extracting significant corner points and curvature mutation points on the edge; using KLT optical flow method to track dynamic displacement of the feature points between continuous video frames; Step S4: applying Fourier transform to displacement data obtained by displacement tracking to convert into frequency domain signal to identify cable fundamental frequency, and using a machine learning model to calculate cable tension based on the fundamental frequency and related parameters; The double thresholding in S2 employs an adaptive thresholding method, including calculating the mean value of the gradient amplitude in the local region and the standard deviation , and dynamically calculating the adaptive high threshold value , low threshold value according to the formula ； ； wherein, is an adjustment parameter; mean represents the overall gradient level of the region; standard deviation represents the gradient volatility; the adjustment parameter The optimal value is found in the preset range [0.5, 2.5] by dichotomy. The edge connection is realized by the following method: taking strong edge pixels obtained after double thresholding as starting point, checking weak edge pixels in the eight neighborhoods thereof; if the weak edge pixels are connected with any strong edge pixel, the weak edge pixels are promoted to strong edge pixels, otherwise, the weak edge pixels are set to zero to obtain continuous edge image.

2. The edge recognition based non-target cable force identification method according to claim 1, characterized in that, The image smoothing in S2 adopts Gaussian filtering, and the gradient calculation uses Sobel operator to calculate image horizontal direction gradient component and vertical direction gradient component.

3. The method of claim 1, wherein, The non-maximum suppression in S2 is realized by the following method: after calculating gradient amplitude and direction, comparing gradient amplitude of each pixel with gradient amplitudes of adjacent two pixels in the gradient direction of the pixel; If it is not a local maximum value, the gradient amplitude of the pixel is set to zero.

4. The method of claim 1, wherein the method is a non-targeting cable force identification method based on edge recognition. The KLT optical flow method in S3 is based on brightness constant assumption, small motion displacement assumption and spatial consistency assumption.

5. The edge recognition based non-target cable force identification method of claim 4, wherein, Based on the spatial consistency assumption, the displacement vector of the feature point is calculated by solving the optical flow equation by least squares method ; wherein, , is an optical flow vector, collectively representing a motion vector of the cable edge feature point between adjacent frames, denotes a horizontal component of the motion vector, denotes a vertical component of the motion vector.

6. The edge recognition based non-target cable force identification method of claim 4, wherein, The KLT optical flow method in S3 adopts image pyramid structure, the pyramid reduces resolution layer by layer from bottom layer original image to top layer with a descending sampling factor, which is used to capture large displacement in coarse resolution layer and refine to fine resolution layer.

7. The method of claim 1, wherein the method is a non-targeting cable force identification method based on edge recognition. The fundamental frequency identification in S4 comprises identifying the main vibration frequency with the largest amplitude in the frequency domain signal as the fundamental frequency.

8. The edge recognition based non-target cable force identification method of claim 1, wherein, The machine learning model in S4 is an XGBOOST regression model, and the related parameters comprise cable length and linear density.

Images