Machine vision-based method for detecting deformation of anti-buoyancy anchor bolts

Abstract

This invention relates to the field of building foundation pit deformation detection technology, specifically a machine vision-based method for detecting the deformation of anti-buoyancy anchor bolts. The method includes setting a high-contrast array of artificial marker points on the anchor bolt observation section, and using a fixed industrial camera to periodically acquire image sequences. An improved template matching algorithm is used to locate the pixel coordinates of the marker points, which are compared with a reference image to obtain a two-dimensional pixel displacement vector. This vector is then converted into a local three-dimensional spatial displacement by combining the camera's intrinsic and extrinsic parameters, constructing a three-dimensional displacement field. This three-dimensional displacement field drives a parameterized elastic foundation beam flexural deformation model to solve for and invert the axial strain distribution and lateral deflection curves of the anchor bolt. This method achieves non-contact, full-domain three-dimensional deformation acquisition. By coupling the three-dimensional displacement field with the foundation beam model, axial strain and lateral deflection can be acquired simultaneously, improving the completeness and accuracy of anti-buoyancy anchor bolt deformation detection.

Description

Technical Field

[0001] This invention relates to the field of building foundation pit deformation detection technology, and in particular to a machine vision-based method for detecting deformation of anti-buoyancy anchor bolts. Background Technology

[0002] In the field of building construction, anti-buoyancy anchors are key components for ensuring the stability of buildings, and accurate monitoring of their deformation is crucial. Currently, there are two main methods for detecting the deformation of anti-buoyancy anchors: contact sensor monitoring and conventional machine vision inspection.

[0003] Contact monitoring methods primarily rely on equipment such as strain gauges and displacement meters to collect deformation data at local points. In practice, the installation process for these contact sensors is extremely complex. Construction sites are typically harsh environments, filled with various construction materials, equipment, and complex wiring. This makes sensor installation not only require a significant amount of manpower and time but also highly susceptible to interference from factors such as mechanical vibration, humidity, and dust at the construction site, thus affecting the accuracy of data acquisition. Furthermore, since contact sensors can only be fixed at a limited number of discrete points, their measurement coverage is very limited. They can only acquire deformation information at these discrete points and cannot comprehensively and continuously reflect the overall deformation of the anti-buoyancy anchor.

[0004] Conventional machine vision inspection methods rely on the natural features of anchor bolts or single marker points, using basic template matching algorithms to extract image features. However, this approach has several limitations. Firstly, conventional machine vision can only extract two-dimensional pixel displacement information from marker points. Because the internal and external parameters of the industrial camera are not calibrated, various distortions during the camera imaging process cannot be corrected, resulting in an inability to accurately map pixel displacement to the real space. Consequently, it cannot generate three-dimensional spatial displacement data of the anchor bolt surface, making it difficult to comprehensively and accurately describe the deformation state of the anchor bolt in actual space. Secondly, the displacement data collected by this method does not undergo in-depth mechanical modeling and analysis in conjunction with foundation constraints, remaining only at the level of simple deflection calculation. This makes it impossible to distinguish between axial deformation and lateral bending deformation of the anchor bolt when facing complex engineering conditions, unable to obtain continuous axial strain distribution, and therefore unable to reconstruct a complete lateral deflection curve. The analysis of anchor bolt deformation lacks depth and comprehensiveness, failing to meet the needs of accurate anchor bolt deformation assessment in practical engineering. In summary, both existing methods for detecting anchor bolt deformation have certain shortcomings, and a more advanced, comprehensive, and accurate detection method is urgently needed to solve these problems. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a machine vision-based method for detecting deformation of anti-buoyancy anchor bolts.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a machine vision-based method for detecting deformation of anti-buoyancy anchor bolts, specifically including the following steps: S1. Establish a high-contrast array of artificial marker points on the surface of the pre-defined observation section of the anchor bolt; S2. Using a fixed industrial camera, continuously acquire image sequences of the anchor bolt area containing the high-contrast artificial marker array at set time intervals; S3. For each frame of the image sequence of the anchor bolt area, use an improved template matching algorithm to identify and locate the pixel coordinates of all high-contrast artificial markers in the image; S4. Compare the pixel coordinates of the high-contrast artificial markers identified in each frame of the image with the pixel coordinates of the corresponding high-contrast artificial markers in the reference image, and calculate the two-dimensional pixel displacement vector of each high-contrast artificial marker. S5. Based on the pre-calibrated internal and external parameters of the industrial camera, the two-dimensional pixel displacement vector is converted to the local three-dimensional space of the anchor bolt surface to obtain the three-dimensional spatial displacement of each high-contrast artificial marker point. S6. Based on the three-dimensional spatial displacement of each high-contrast artificial marker point, construct a three-dimensional displacement field on the surface of the anchor observation section; S7. Using the three-dimensional displacement field on the surface of the anchor observation section, drive a parameterized elastic foundation beam flexural deformation analysis model; S8. By solving the parameterized elastic foundation beam flexural deformation analysis model, the axial strain distribution and lateral deflection curve of the anchor rod are obtained.

[0007] In this embodiment, step S1, establishing a high-contrast artificial marker array on the surface of the preset observation section of the anchor bolt, specifically includes: S11. Spray or paste circular or ring-shaped markers with specific reflectance spectral characteristics on the surface of the observation section of the anchor bolt; S12. Adjust the angle and intensity of the illumination source of the industrial camera to ensure that the marker and the anchor background form the maximum grayscale or color contrast in the acquired image. S13. When the anchor bolt is in an initial state without load, the industrial camera captures a clear image as the reference image. S14. In the reference image, manually or automatically select the area where all markers are located; S15. Perform ellipse fitting on each marker region and calculate its centroid coordinates. The centroid coordinates are used as the pixel coordinates of the high-contrast artificial markers in the reference image.

[0008] In this embodiment, step S2, while acquiring image sequences through the industrial camera, also includes an environmental interference suppression step: real-time monitoring of the average brightness and contrast of the acquired images; when the average brightness is lower than a set threshold, adjusting the intensity of the supplementary light attached to the industrial camera or the exposure time of the industrial camera; when the image contrast is lower than a set threshold, using a limited contrast adaptive histogram equalization method to preprocess the acquired original image before proceeding to the subsequent marker recognition process.

[0009] In this embodiment, the method for setting the time interval in step S2 includes: using a short time interval for high-frequency image acquisition in the initial stage of anchor bolt loading or in the stage of drastic load change; using a long time interval for low-frequency image acquisition in the stage of anchor bolt deformation stabilization; receiving an externally input load change threshold; and switching from low-frequency acquisition mode to high-frequency acquisition mode when the monitoring system infers that the anchor bolt load change exceeds the load change threshold.

[0010] In this embodiment, step S2 requires pre-calibrating the internal and external parameters of the industrial camera. The specific steps are as follows: a stereo calibration plate with known precise three-dimensional dimensions is installed on a fixed structure near the anchor bolt; the industrial camera is controlled to capture images containing the stereo calibration plate from multiple different angles to form a calibration image set; the image coordinates of feature points on the stereo calibration plate are detected from the calibration image set; combined with the known world coordinates of the feature points on the stereo calibration plate, the intrinsic parameter matrix, distortion coefficient, and extrinsic parameter matrix of the industrial camera relative to the coordinate system of the stereo calibration plate are solved by minimizing the reprojection error; and a fixed transformation relationship is established between the coordinate system of the industrial camera and the local three-dimensional coordinate system of the anchor bolt through coordinate system transformation.

[0011] In this embodiment, the working principle of the improved template matching algorithm in step S3 includes: S31. In the reference image, take the center of each high-contrast artificial marker point as the origin and extract a standard-sized image block as the initial template. S32. In subsequent images to be detected, a search area larger than the initial template is defined, centered on the theoretically predicted position of the high-contrast artificial marker. S33. Within the search area, normalized cross-correlation is used to calculate the similarity between the initial template and each sub-region of the image to be detected, and a similarity distribution map is obtained. S34. In the similarity distribution map, locate the similarity peak point and preliminarily determine the candidate positions of high-contrast artificial marker points; S35. Extract the image patch centered on the candidate position, and input it together with the initial template into a lightweight convolutional neural network for feature re-extraction to obtain a deep feature descriptor; S36. Calculate the cosine similarity between the deep feature descriptors. When the cosine similarity is higher than a set threshold, confirm the candidate position as the pixel position of the high-contrast artificial marker.

[0012] In this embodiment, step S5, converting the two-dimensional pixel displacement vector to the local three-dimensional space of the anchor surface, includes: S51. Using the camera intrinsic parameter matrix calibrated by the industrial camera in step 2, eliminate the influence of image distortion on pixel coordinates and obtain the normalized camera coordinates of each high-contrast artificial marker point. S52. Based on the extrinsic parameter matrix of the industrial camera relative to the coordinate system of the anchor bolt observation section, transform the normalized camera coordinates to the local three-dimensional coordinate system of the anchor bolt surface; S53. In the local three-dimensional coordinate system, the three-dimensional coordinates of each high-contrast artificial marker point in the reference image are known, and the three-dimensional coordinates of the corresponding point in the current frame image are obtained through the aforementioned coordinate transformation. S54. Calculate the coordinate difference between the same high-contrast artificial marker point in the current frame and the reference image in the local three-dimensional coordinate system. The coordinate difference is the three-dimensional spatial displacement of the high-contrast artificial marker point.

[0013] In this embodiment, step S6, constructing the three-dimensional displacement field on the surface of the anchor observation section, includes: S61. Using the axial direction of the anchor bolt as the reference direction, the observation section is divided into several continuous micro-segments in the axial direction; S62. Decompose the three-dimensional spatial displacement of all high-contrast artificial markers located in each micro-segment into three components: axial, radial and circumferential. S63. Spatial interpolation is performed on the surface of the observation section for the axial displacement component, radial displacement component and circumferential displacement component respectively to generate three independent and continuous displacement component surfaces. S64. The three independent and continuous displacement component surfaces are synthesized to form a vector field that characterizes the three-dimensional motion information of each point on the surface of the observation segment, namely the three-dimensional displacement field.

[0014] In this embodiment, step S7, the driving and solving of the parameterized elastic foundation beam flexural deformation analysis model includes: abstracting the anchor rod as an Euler-Bernoulli beam located in an elastic medium; using the radial displacement component surface in the three-dimensional displacement field as the boundary condition for the measured deflection distribution of the Euler-Bernoulli beam; using the axial displacement component surface in the three-dimensional displacement field to derive the strain distribution along the anchor rod axis as the constraint condition for the axial force distribution; setting the elastic modulus, moment of inertia of the section, and foundation reaction coefficient of the anchor rod as the parameters to be identified in the model; using the measured deflection distribution boundary condition and the axial force distribution constraint condition as the target, adjusting the parameters to be identified in the model through an iterative optimization algorithm; stopping the iteration when the error between the deflection curve and axial strain distribution calculated by the model and the measured value derived from the three-dimensional displacement field is the smallest, and the model state at this time is the inversion result.

[0015] In this embodiment, after inverting the axial strain distribution and lateral deflection curve of the anchor rod in step S7, a deformation state assessment step is also included: extracting the axial strain distribution curve along the entire length of the anchor rod from the final inversion result of the parameterized elastic foundation beam flexural deformation analysis model; extracting the lateral deflection curve along the entire length of the anchor rod from the final inversion result; calculating the first derivative of the lateral deflection curve to obtain the angle change curve of the anchor rod; calculating the second derivative of the lateral deflection curve, and deriving the bending moment distribution curve of the anchor rod section by combining the elastic modulus information of the anchor rod.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. This method utilizes a high-contrast artificial marker array to improve the feature discrimination of the anchor bolt observation section. An improved template matching algorithm can stably identify the marker positions in the image and accurately extract the pixel coordinates of each marker point frame by frame. By comparing the real-time frame marker coordinates with the reference image coordinates, a two-dimensional pixel displacement vector for a single marker point can be obtained. Relying on the pre-calibrated internal and external parameters of the industrial camera, imaging system errors and spatial perspective deviations can be eliminated, converting the two-dimensional pixel displacement vector into a local three-dimensional spatial displacement on the anchor bolt surface. The combination of multiple marker point three-dimensional spatial displacements can construct a global three-dimensional displacement field for the anchor bolt observation section, fully presenting the spatial deformation state at various locations on the anchor bolt surface. This overcomes the limitations of traditional local single-point detection and two-dimensional plane detection, achieving a quantitative representation of global spatial deformation.

[0018] 2. The three-dimensional displacement field of this method can be directly used as input data to drive the parameterized elastic foundation beam flexural deformation analysis model, making the model loading conditions fit the actual spatial deformation shape of the anchor rod. The model solution process can be coupled with the elastic foundation constraint, independently distinguishing the axial tensile deformation and lateral bending deformation components. Through numerical calculation of the model, the axial strain distribution that changes continuously along the anchor rod length can be obtained, and a complete and smooth lateral deflection curve can be generated. The combination of displacement field data and mechanical analysis model can realize the transformation from geometric deformation to mechanical parameters, refine the deformation detection dimension, obtain continuously distributed strain parameters and complete deflection shape, improve the detail and characterization dimension of deformation solution, and realize the synchronous solution and quantitative presentation of multiple types of deformation parameters. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention; Figure 2 A flowchart for establishing a high-contrast array of artificial marker points on the surface of an anchor bolt; Figure 3 A flowchart illustrating the work done on the improved template matching algorithm. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] like Figure 1 This invention provides a machine vision-based method for detecting deformation of anti-buoyancy anchor bolts, the overall implementation of which is as follows:

[0022] On the surface of the pre-defined observation section of the anti-buoyancy anchor bolt requiring deformation monitoring, a high-contrast array of artificial marker points is established. An industrial camera is fixed in a stable position capable of completely capturing the observation section of the anchor bolt, and the image acquisition time interval is set according to the monitoring requirements, thereby continuously acquiring an image sequence containing the marker point array during the anchor bolt's stress process. For each frame of the sequence, from 30° to 45°, an improved template matching algorithm is used to identify all high-contrast artificial marker points in the image and obtain their pixel coordinates in the image. The pixel coordinates of the marker points identified in each frame are compared with the pixel coordinates of the corresponding marker points in a pre-selected reference image. The markers are compared, and the two-dimensional pixel displacement vector of each marker point on the image plane is calculated point by point. Based on the internal and external parameters of the industrial camera that have been calibrated in advance, the two-dimensional pixel displacement vector is converted into the local three-dimensional spatial coordinate system of the anchor bolt surface to obtain the three-dimensional spatial displacement of each marker point. Based on the three-dimensional spatial displacement data of all marker points on the surface of the observation section, a continuous three-dimensional displacement field reflecting the overall motion state of the anchor bolt surface is constructed. This three-dimensional displacement field is used as the driving condition and input into a parameterized elastic foundation beam flexural deformation analysis model. By numerically solving the model, the true strain distribution of the anchor bolt along the axial direction and its lateral deflection curve are obtained.

[0023] In one embodiment of the present invention, such as Figure 2 On the surface of the pre-defined observation section exposed by the anchor bolt, a series of circular or ring-shaped markers arranged in an array are set by spraying or pasting. The selected markers have specific reflectance spectral characteristics that are different from the anchor bolt body material. The illumination angle and output intensity of the industrial camera's lighting source are adjusted to maximize the grayscale difference or color distinction between the marker area and the surrounding anchor bolt background in the acquired image. In the initial state where the anchor bolt has not yet been subjected to external load, a clear image is captured by the industrial camera and set as the reference image. In this reference image, the image area where all markers are located is defined by manual selection or automatic recognition program. An ellipse fitting operation is performed on each selected marker area, and the calculated ellipse geometric centroid coordinates are recorded as the reference pixel coordinates of the high-contrast artificial marker point in the reference image.

[0024] In practical implementation, for the exposed surface of the anti-buoyancy anchor bolt designated as the observation section, white high-reflectivity paint is sprayed using a positioning mold to form circular markers of uniform diameter. These circular markers are arranged in three rows at equal intervals along the anchor bolt axis, with adjacent rows staggered to form an artificial marker array. This ensures that any adjacent high-contrast artificial markers in the image are clearly distinguishable without overlap. In practice, an adjustable-angle LED strip cold light source is installed with the industrial camera. The incident angle of the LED strip cold light source is adjusted to be between 30° and 45° with the normal to the anchor bolt surface. Simultaneously, the output current of the LED strip cold light source is adjusted to control the light intensity until the difference between the average grayscale value of the high-contrast artificial marker area and the average grayscale value of the black oxidized background of the anchor bolt in the acquired preview image reaches its maximum. This light source configuration is then locked to ensure maximum grayscale contrast between the markers and the anchor bolt background in subsequent acquired images.

[0025] In some embodiments, when the anchor bolt is in its initial unloaded state after construction is completed and no engineering load is applied, an industrial camera is triggered to acquire a clear, normally exposed image at the highest resolution. This image is stored and designated as the reference image for the entire deformation detection process. On the reference image display interface, the operator manually drags a rectangular selection box to sequentially select the pixel regions containing all high-contrast artificial markers in the image. Each rectangular selection box completely surrounds a single high-contrast artificial marker with a safety margin of at least 5 pixels at the edges. It is understood that the automatic recognition program can also scan the reference image using a preset speckle detection algorithm, filter out all potential high-contrast artificial marker regions based on a connected component area threshold, and automatically generate a selection area.

[0026] In practice, an ellipse fitting operation is performed on each selected high-contrast artificial marker region. The coordinates of all foreground pixels within the region are fitted to the best approximating ellipse. The objective function of the ellipse fitting is defined as minimizing the sum of squared algebraic distances from all foreground pixels to the ellipse boundary. The center coordinates, major and minor axes, and rotation angle parameters of the ellipse are obtained by solving the normal equation. Optionally, the ellipse fitting process can employ a random sampling consensus algorithm to remove outlier pixels caused by image noise, thereby improving the robustness of centroid localization. For any high-contrast artificial marker region, the center coordinates of the fitted ellipse are... The calculation expression is: ; in: This represents the ordinate and abscissa values ​​of the ellipse's center in the image pixel coordinate system. Represents the first ellipse on the fitted elliptical trajectory The pixel coordinates of each sample point This represents the total number of elliptical trajectory sample points involved in the centroid calculation. It can be understood that these centroid coordinates serve as the reference pixel coordinates of the corresponding high-contrast artificial marker point in the reference image. The reference pixel coordinates of all high-contrast artificial marker points are aggregated to form the reference benchmark for subsequent displacement calculations. In some embodiments, if the high-contrast artificial marker point exhibits a non-circular elliptical feature in the image due to perspective projection, the fitted elliptical centroid coordinates are still considered as the representative position of the marker point. Its systematic deviation will be compensated for through perspective correction during subsequent coordinate system calibration and transformation.

[0027] In one embodiment of the present invention, such as Figure 3 Within the baseline image, a fixed-size square image patch is extracted with the centroid of each high-contrast artificial marker as the origin, and saved as the initial template for subsequent matching. When processing subsequently acquired images to be detected, the approximate position of each marker is estimated based on the possible deformation trend of the anchor rod, and a rectangular search area much larger than the initial template size is delineated centered on this. Within this search area, a normalized cross-correlation calculation method is used to traverse every possible position to obtain the similarity score between the initial template and the corresponding sub-region of the image to be detected, thereby generating a similarity distribution map of the entire search area. The global maximum point is found in the similarity distribution map, and its coordinates are used as the coarse localization candidate position of the marker. An image patch is extracted around this candidate position, and it and the initial template are input into a pre-trained lightweight convolutional neural network. The network output is the deep feature descriptor of the two. The cosine similarity between the two descriptors is calculated. If the calculation result is higher than the preset judgment threshold, the candidate position is confirmed as valid, which is the precise pixel position of the high-contrast artificial marker in the current frame. Before performing the above image processing, the coordinate system needs to be calibrated. Specifically, this involves: installing a three-dimensional calibration plate with known precise dimensions on a fixed structure near the anchor; using an industrial camera to take multiple photos containing this calibration plate from different perspectives to form a calibration image set; extracting the pixel coordinates of regularly arranged feature points on the calibration plate from the calibration images; combining the known coordinate values ​​of these feature points in the world coordinate system, using a nonlinear optimization algorithm that minimizes reprojection error to calculate the intrinsic parameter matrix and distortion coefficients of the industrial camera, as well as the extrinsic parameter matrix of the camera relative to the calibration plate coordinate system; and establishing the mapping relationship between the industrial camera coordinate system and the local three-dimensional coordinate system of the anchor through a fixed coordinate transformation chain.

[0028] In the specific implementation, in the selected reference image, a square image block with a side length of 32 pixels is extracted, centered on the reference pixel coordinates of each high-contrast artificial marker point in the reference image. This square image block is used as the initial template for subsequent image matching and stored in the template library. In the specific implementation, for each high-contrast artificial marker point in the image to be detected, based on the assumption of anchor deformation continuity, the theoretical predicted position is extrapolated along the deformation trend based on the pixel coordinates of the high-contrast artificial marker point in the previous frame image. A rectangular search area with a side length of 80 pixels is delineated centered on the theoretical predicted position. This rectangular search area completely includes any possible translation range of the initial template.

[0029] Within the rectangular search area, the initial template is slid pixel by pixel and the corresponding sub-region of the image to be detected is used to calculate the normalized cross-correlation score. The formula for calculating the normalized cross-correlation is: ; in: This indicates that the initial template and the image to be detected are offset. Normalized cross-correlation coefficient at the point, Indicates the initial template in local coordinates The pixel grayscale value at that location, This represents the average grayscale value of all pixels in the initial template. Indicates the image to be detected in global coordinates The pixel grayscale value at that location, Indicates that in the image to be detected, The grayscale mean of pixels in the top-left sub-region, which is the same size as the initial template. After calculating the normalized cross-correlation distribution map of the entire rectangular search area, the coordinates of the maximum peak point in the distribution map are used as candidate positions for high-contrast artificial markers.

[0030] Optionally, in certain scenarios with fluctuating illumination, the peak value of the normalized cross-correlation may be obscured by local maxima. In such cases, the sub-pixel offset of the candidate location's neighborhood can be further calculated to correct the candidate location coordinates. It is understood that the normalized cross-correlation method is invariant to linear changes in illumination, making it suitable for the slow-changing natural illumination conditions that may exist at the anchor bolt site.

[0031] In specific implementation, a 32×32 pixel image patch centered on the candidate location of a high-contrast artificial marker in the image to be detected is extracted. This image patch, along with the corresponding initial template, is input into a lightweight convolutional neural network (CNN). The CNN contains three convolutional layers and two fully connected layers, outputting a 128-dimensional deep feature descriptor. The cosine similarity between the deep feature descriptor of the image patch to be detected and the deep feature descriptor of the initial template is calculated. When the cosine similarity is greater than a set threshold of 0.95, the candidate location is confirmed as valid; otherwise, the high-contrast artificial marker is determined to be occluded or lost. In some embodiments, the CNN is pre-trained on a synthetic marker dataset that includes rotation, scaling, and brightness variations to improve robustness to changes in viewpoint and illumination. Optionally, the cosine similarity threshold can be dynamically adjusted according to the environment, appropriately reduced to 0.90 in rainy or foggy conditions to tolerate image quality degradation.

[0032] In some embodiments, before performing the above image processing, industrial camera parameter calibration needs to be completed. A black and white checkerboard 3D calibration plate with a side length of 300 mm is installed on the side wall of an adjacent fixed concrete pier approximately 1.5 meters from the anchor bolt. The checkerboard has 12×9 grids, and the physical size accuracy of the grid points is ±0.01 mm. The industrial camera is controlled to capture an image containing the 3D calibration plate every 5° from a pitch angle of -20° to +20°, for a total of 9 frames to form a calibration image set. From the calibration image set, the checkerboard corner detection algorithm is used to extract the image pixel coordinates of all interior corner points on the 3D calibration plate. Combined with the known world coordinates in the coordinate system of the 3D calibration plate, the Levenberg-Marquardt nonlinear optimization algorithm is used to minimize the reprojection error, and the intrinsic parameter matrix, radial and tangential distortion coefficients, rotation matrix, and translation vector of the industrial camera relative to the coordinate system of the 3D calibration plate are solved. It is understandable that by using the pre-measured fixed rigid body transformation relationship between the three-dimensional calibration plate coordinate system and the local three-dimensional coordinate system of the anchor bolt, the external parameters of the industrial camera are uniformly transformed to the local three-dimensional coordinate system of the anchor bolt, thus establishing a complete mapping chain from the industrial camera pixel coordinate system to the local three-dimensional coordinate system of the anchor bolt.

[0033] In one embodiment of the present invention, the original pixel coordinates of the identified high-contrast artificial markers are corrected for distortion using the calibrated intrinsic parameter matrix and distortion coefficients of the industrial camera, and then converted into normalized camera coordinates. Based on the rotation matrix and translation vector determined by the extrinsic parameter matrix of the industrial camera relative to the anchor bolt observation section coordinate system, the normalized camera coordinates are transformed into a local three-dimensional coordinate system with the anchor bolt axis as the reference. In this local three-dimensional coordinate system, the three-dimensional coordinates of each marker in the reference image are known fixed values, and the three-dimensional coordinates of the same marker in the current frame image are obtained through the same coordinate transformation process. The differences between the current frame coordinates and the reference frame coordinates in the three-axis directions are calculated one by one, and the result is the actual displacement of the high-contrast artificial marker in the local three-dimensional space. Subsequently, a three-dimensional displacement field was constructed, and the anchor observation section was uniformly divided into numerous continuous micro-unit segments along its axial direction. All the marker points contained in each micro-unit segment were classified, and their three-dimensional displacements were decomposed into axial components parallel to the axis, radial components perpendicular to the axis pointing outward, and circumferential components tangential to the axis. On the surface of the observation section, the axial displacement components of each marker point were spatially interpolated to generate continuous axial displacement surfaces, the radial components were interpolated to generate radial displacement surfaces, and the circumferential components were interpolated to generate circumferential displacement surfaces. These three independent displacement component surfaces were combined into one according to the principle of vector superposition to form a continuous three-dimensional vector field covering the entire surface of the observation section.

[0034] In practical implementation, based on the intrinsic parameter matrix containing focal length and principal point coordinates, and the distortion coefficient vector containing radial distortion coefficients obtained from industrial camera calibration, the original pixel coordinates of the identified high-contrast artificial markers are distorted. A Brown-Conrad model is applied to compensate for radial and tangential distortion. The corrected pixel coordinates are then projected onto the normalized imaging plane through inverse operation of the intrinsic parameter matrix to obtain normalized camera coordinates. In practical implementation, based on the extrinsic parameter matrix of the industrial camera relative to the anchor bolt observation section coordinate system (i.e., the rotation matrix and translation vector), the normalized camera coordinates undergo a homogeneous coordinate transformation, first rotating and then translating to a local three-dimensional coordinate system of the anchor bolt with the anchor bolt axis as the Z-axis and the center of the near-end face of the anchor bolt as the origin. In the local three-dimensional coordinate system of the anchor bolt, the three-dimensional coordinates of each high-contrast artificial marker point in the reference image have been determined and stored during the initial calibration stage. The three-dimensional coordinates of the same high-contrast artificial marker point in the current frame image are calculated in real time through the same coordinate transformation process. The difference between the X, Y, and Z axis coordinates of the same high-contrast artificial marker point in the current frame and the reference image in the local three-dimensional coordinate system of the anchor bolt is calculated. This difference is the three-dimensional spatial displacement of the high-contrast artificial marker point. Refer to Table 1 for the calculation results of the local three-dimensional coordinates and displacements of a set of high-contrast artificial marker points in the reference state and the current frame:

[0035] Table 1: Results of Local Three-Dimensional Coordinate and Displacement Calculation ; It is understandable that the calculation of coordinate differences follows strict algebraic operation rules, including axial displacement components. A negative value indicates that the anchor bolt undergoes axial compression deformation at the high-contrast artificially marked point.

[0036] In the specific implementation, when constructing the three-dimensional displacement field on the surface of the anchor bolt observation section, the observation section is uniformly divided into continuous cylindrical micro-segments with a length of 50 mm along the anchor bolt axis. Each micro-segment contains at least 3 high-contrast artificial marker points. The three-dimensional spatial displacement of all high-contrast artificial marker points located in each micro-segment is decomposed into axial, radial, and circumferential components according to the definition of the local three-dimensional coordinate system of the anchor bolt. The axial component is parallel to the Z-axis, the radial component is perpendicular to the Z-axis and points away from the axis, and the circumferential component is perpendicular to the Z-axis and orthogonal to the radial component. For the axial displacement component, a bicubic spline interpolation method is used to interpolate on the grid nodes of the observation section surface to generate a continuous axial displacement component surface. The interpolation weight is determined by the square of the inverse distance from the high-contrast artificial marker point to the grid node. The interpolation formula is expressed as: ; in: For grid nodes The axial displacement interpolation results at the location, This represents the number of high-contrast artificial markers within the current micro-segment. For the first Axial displacement components of high-contrast artificially marked points This represents the interpolation weight of the k-th high-contrast artificial marker point. For the first The spatial coordinates of a high-contrast artificially marked point. For Euclidean distance, To prevent small constants with zero denominators, similarly, the same interpolation process is performed on the radial and circumferential displacement components to generate continuous radial and circumferential displacement component surfaces. In some embodiments, when the number of high-contrast artificial markers within a certain micro-element is insufficient, the interpolation neighborhood is automatically expanded to the set of high-contrast artificial markers in adjacent micro-elements to ensure surface smoothness. Optionally, thin-plate spline interpolation can be used instead of bicubic spline interpolation to adapt to non-uniformly distributed arrays of high-contrast artificial markers. It can be understood that the synthesis of three independent displacement component surfaces forms a three-dimensional vector field covering the entire surface of the observation segment, and the vector at each grid point in the vector field is composed of the axial, radial, and circumferential displacement components corresponding to that point.

[0037] In one embodiment of the present invention, the actual anti-buoyancy anchor is abstracted as an Euler Bernoulli beam model placed in a Winkler elastic foundation; the radial displacement component surface is extracted from the constructed three-dimensional displacement field, and its distribution law along the anchor axis is used as the measured deflection boundary condition of the beam model; the axial displacement component surface is extracted from the three-dimensional displacement field, and the axial strain distribution of the anchor is derived by its axial gradient, which is used as the equilibrium constraint condition of the axial force of the model; the elastic modulus, cross-sectional moment of inertia and foundation reaction coefficient of the anchor material are set as unknown parameters that need to be obtained through inverse analysis; the measured deflection distribution boundary condition and axial strain constraint condition are used as fitting targets, and the iterative optimization algorithm is run to continuously adjust the values ​​of the above unknown parameters; the iteration is terminated when the total residual between the beam deflection curve obtained by the forward calculation of the model and the axial strain distribution and the measured data derived from the three-dimensional displacement field reaches the minimum value, and the convergence state of the model at this time is the optimal solution of the inversion. The axial strain values ​​distributed along the length of the anchor rod are read from the optimal solution to form the axial strain distribution curve; the lateral displacement values ​​along the length of the anchor rod calculated by the model are output simultaneously to form the lateral deflection curve; the first derivative operation of the lateral deflection curve is performed to obtain the rotation angle change curve of each point of the anchor rod; the second derivative operation of the lateral deflection curve is performed, and the result is multiplied by the elastic modulus and the moment of inertia of the anchor rod to derive the bending moment distribution curve along the entire length of the anchor rod.

[0038] In practical implementation, the anti-buoyancy anchor in actual engineering is abstracted as an Euler-Bernoulli beam embedded in a Winkler-type elastic medium. The length of this beam is equal to the length of the anchor observation section, and the bending stiffness and axial stiffness of the beam correspond to the elastic modulus and cross-sectional geometric properties of the anchor material, respectively. In the specific implementation, a radial displacement component surface is extracted from the constructed three-dimensional displacement field. A set of radial displacement values ​​is obtained by discrete sampling at 100 mm intervals along the anchor axis. This set of radial displacement values ​​is used as the measured deflection boundary condition of the Euler-Bernoulli beam at the corresponding positions. An axial displacement component surface is extracted from the three-dimensional displacement field. The axial displacement difference between adjacent sampling points is calculated along the anchor axis and divided by the sampling interval to derive the axial strain distribution along the anchor axis. This axial strain distribution is used as the constraint condition for the axial force balance of the Euler-Bernoulli beam. The elastic modulus of the anchor material, the moment of inertia of the anchor cross-section, and the soil reaction coefficient are set as the parameters to be identified in the model. A real-number encoded genetic algorithm is used as the iterative optimization algorithm. In each iteration, the values ​​of the parameters to be identified in the model are adjusted and substituted into the governing differential equation of the Euler-Bernoulli beam to calculate the deflection curve and axial strain distribution of the beam. See Table 2 for a schematic record of the adjustment of the parameters to be identified and the changes in the residuals during one iteration.

[0039] Table 2: A schematic record of the adjustment of the model's parameters to be identified and the changes in the residuals during one iteration. ; It can be understood that the deflection residual norm is the L2 norm of the difference between the model-calculated deflection curve and the measured deflection curve derived from the three-dimensional displacement field, and the strain residual norm is the L2 norm of the difference between the model-calculated axial strain distribution and the measured axial strain distribution. Iteration stops when the weighted sum of the deflection residual norm and the strain residual norm is less than the convergence threshold; at this point, the model's state is the optimal solution obtained through inversion.

[0040] The axial strain values ​​discretely distributed along the entire length of the anchor rod are extracted from the final inversion results, and the axial strain distribution curve of the anchor rod is plotted. The lateral displacement values ​​along the entire length of the anchor rod are also extracted from the final inversion results, and the lateral deflection curve of the anchor rod is plotted. The first-order derivative of the lateral deflection curve is performed, with the difference step size being the sampling interval along the anchor rod axis, to obtain the rotation angle variation curves of each discrete node of the anchor rod. The second-order derivative of the lateral deflection curve is performed, and combined with the elastic modulus and section moment of inertia obtained from the inversion, the bending moment distribution curve of the anchor rod section is calculated based on the Euler Bernoulli beam bending moment-curvature relationship. The bending moment calculation formula is: ; in: Indicates the anchor rod in arc coordinates Bending moment at section, This represents the elastic modulus of the anchor material obtained through inversion. This represents the moment of inertia of the anchor bolt cross section obtained through inversion. Indicates the lateral deflection curve at The approximate value of the second derivative at the given point. In some embodiments, to improve numerical stability, the second derivative is calculated using a five-point central difference scheme. Optionally, when the monitoring focus is on the bending and shear performance of the anchor bolt, the shear force distribution curve can be obtained by further performing a first-order difference derivative on the bending moment distribution curve.

[0041] In one embodiment of the present invention, during the timed image acquisition task performed by the industrial camera, an environmental monitoring module runs in parallel to statistically analyze the average brightness and local contrast index of each captured image frame in real time. When the average brightness of the image is continuously lower than the safety threshold, a feedback signal controls the intensity of the supplementary light connected to the industrial camera to increase or the exposure time of the single frame of the industrial camera to be extended. When the image contrast is detected to drop to a set lower limit due to fog or lens damage, the image preprocessing module is immediately started, and the original image is enhanced and sharpened using a contrast-limited adaptive histogram equalization algorithm. The processed image is then sent to the marker point recognition process. Regarding the setting of the image acquisition time interval, a shorter time interval is used to perform high-frequency dense image capture when the anchor bolt is just beginning to be loaded or when there is a significant change in the external load. When the monitoring data shows that the deformation rate of the anchor bolt slows down and the shape tends to be stable, the system automatically switches to a longer time interval to perform low-frequency energy-saving acquisition. The system receives a load fluctuation warning value input from an external monitoring platform. Once the logic determines that the load change amplitude of the anchor bolt exceeds the threshold, the acquisition mode is immediately switched back from low frequency to high frequency.

[0042] In practical implementation, an environmental monitoring thread runs synchronously during the acquisition of image sequences by the industrial camera. Within this thread, the average grayscale value of all pixels in each frame is calculated in real time as the average image brightness, and the ratio of the image grayscale standard deviation to the brightness is calculated as the contrast index. Specifically, when the average image brightness is below a set threshold of 120 (8-bit grayscale range 0-255), a command is sent via serial port to linearly increase the duty cycle of the industrial camera's attached LED fill light by 5% per frame until the average image brightness rises above 140. Alternatively, if the fill light reaches its maximum power but still cannot meet the brightness requirements, the exposure time of the industrial camera is automatically extended from the default 8 milliseconds to 20 milliseconds. When the image contrast index is below a set threshold of 0.35, the image enhancement processing pipeline is triggered. A contrast-limited adaptive histogram equalization algorithm is applied to the original acquired image, dividing the image into 8×8 sub-blocks for individual equalization. Bilinear interpolation is performed at the sub-block boundaries for smooth transition. After processing, the enhanced image is output to the subsequent high-contrast manual marker recognition module. In some embodiments, the cutoff ratio of adaptive histogram equalization is set to 0.015 to prevent excessive amplification of local noise, and the formula for determining the contrast recovery effect is: ; in: This indicates the contrast quality index of the enhanced image. This represents the global grayscale standard deviation of the image. Indicates the average brightness of the image. The acceptable contrast threshold is 0.38. When this happens, the image will be discarded and reacquired to avoid introducing errors in marker positioning due to poor-quality images.

[0043] In practical implementation, a dual-mode switching strategy is adopted for the time interval setting. In the initial stage of anchor loading or during periods of drastic load changes due to foundation pit excavation, the time interval is set to 200 milliseconds, with the industrial camera continuously acquiring images at a frequency of 5 Hz. When monitoring data shows that the anchor deformation increment is less than 0.05 mm for 10 consecutive minutes, it is determined that the deformation is stabilizing, and the time interval automatically switches to 60,000 milliseconds, switching to a low-frequency mode of acquiring images once per minute. The system receives a load change threshold of 50 kN from an external building structure monitoring platform. When the real-time calculated change in anchor axial force exceeds 50 kN, the acquisition controller immediately switches from low-frequency mode back to high-frequency acquisition mode. Optionally, an anti-jitter delay is set for the time interval switching, triggering mode switching only after the load exceeding the threshold event lasts for more than 2 seconds, avoiding frequent mode oscillations caused by instantaneous disturbances. It can be understood that the high-frequency acquisition mode prioritizes ensuring the spatiotemporal resolution of the transient deformation process, while the low-frequency acquisition mode considers data redundancy and storage costs for long-term monitoring.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A machine vision-based method for detecting deformation of anti-buoyancy anchor bolts, characterized in that, Specifically, the steps include the following: S1. Establish a high-contrast array of artificial marker points on the surface of the pre-defined observation section of the anchor bolt; S2. Using a fixed industrial camera, continuously acquire image sequences of the anchor bolt area containing the high-contrast artificial marker array at set time intervals; S3. For each frame of the image sequence of the anchor bolt area, use an improved template matching algorithm to identify and locate the pixel coordinates of all high-contrast artificial markers in the image; S4. Compare the pixel coordinates of the high-contrast artificial markers identified in each frame of the image with the pixel coordinates of the corresponding high-contrast artificial markers in the reference image, and calculate the two-dimensional pixel displacement vector of each high-contrast artificial marker. S5. Based on the pre-calibrated internal and external parameters of the industrial camera, the two-dimensional pixel displacement vector is converted to the local three-dimensional space of the anchor bolt surface to obtain the three-dimensional spatial displacement of each high-contrast artificial marker point. S6. Based on the three-dimensional spatial displacement of each high-contrast artificial marker point, construct a three-dimensional displacement field on the surface of the anchor bolt observation section; constructing the three-dimensional displacement field on the surface of the anchor bolt observation section includes: S61. Using the axial direction of the anchor bolt as the reference direction, the observation section is divided into several continuous micro-segments in the axial direction; S62. Decompose the three-dimensional spatial displacement of all high-contrast artificial markers located in each micro-segment into three components: axial, radial and circumferential. S63. Spatial interpolation is performed on the surface of the observation section for the axial displacement component, radial displacement component and circumferential displacement component respectively to generate three independent and continuous displacement component surfaces. S64. The three independent and continuous displacement component surfaces are synthesized to form a vector field characterizing the three-dimensional motion information of each point on the surface of the observation segment, namely the three-dimensional displacement field. S7. Using the three-dimensional displacement field on the surface of the anchor observation section, a parameterized elastic foundation beam flexural deformation analysis model is driven. The driving and solving of the parameterized elastic foundation beam flexural deformation analysis model includes: abstracting the anchor as an Euler-Bernoulli beam located in an elastic medium; using the radial displacement component surface in the three-dimensional displacement field as the boundary condition for the measured deflection distribution of the Euler-Bernoulli beam; using the axial displacement component surface in the three-dimensional displacement field to derive the strain distribution along the anchor axis as the constraint condition for the axial force distribution; setting the elastic modulus, cross-sectional moment of inertia, and foundation reaction coefficient of the anchor as the parameters to be identified in the model; using the measured deflection distribution boundary condition and the axial force distribution constraint condition as the objective, adjusting the parameters to be identified in the model through an iterative optimization algorithm; stopping the iteration when the error between the deflection curve and axial strain distribution calculated by the model and the measured value derived from the three-dimensional displacement field is minimized, and the model state at this time is the inversion result. S8. By solving the parameterized elastic foundation beam flexural deformation analysis model, the axial strain distribution and lateral deflection curve of the anchor rod are obtained.

2. The machine vision-based deformation detection method for anti-buoyancy anchor bolts as described in claim 1, characterized in that, In step S1, establishing a high-contrast array of artificial marker points on the surface of the preset observation section of the anchor bolt specifically includes: S11. Spray or paste circular or ring-shaped markers with specific reflectance spectral characteristics on the surface of the observation section of the anchor bolt; S12. Adjust the angle and intensity of the illumination source of the industrial camera to ensure that the marker and the anchor background form the maximum grayscale or color contrast in the acquired image. S13. When the anchor bolt is in an initial state without load, the industrial camera captures a clear image as the reference image; S14. In the reference image, manually or automatically select the area where all markers are located; S15. Perform ellipse fitting on each marker region and calculate its centroid coordinates. The centroid coordinates are used as the pixel coordinates of the high-contrast artificial markers in the reference image.

3. The machine vision-based deformation detection method for anti-buoyancy anchor bolts as described in claim 1, characterized in that, In step S2, while acquiring image sequences through the industrial camera, an environmental interference suppression step is also included: real-time monitoring of the average brightness and contrast of the acquired images; when the average brightness is lower than a set threshold, adjusting the intensity of the supplementary light attached to the industrial camera or the exposure time of the industrial camera; when the image contrast is lower than a set threshold, using a limited contrast adaptive histogram equalization method to preprocess the acquired original image before proceeding to the subsequent marker recognition process.

4. The machine vision-based deformation detection method for anti-buoyancy anchor bolts as described in claim 1, characterized in that, In step S2, the method for setting the time interval includes: using a short time interval for high-frequency image acquisition in the initial stage of anchor bolt loading or in the stage of drastic load change; using a long time interval for low-frequency image acquisition in the stage of anchor bolt deformation stabilization; receiving an externally input load change threshold; and switching from low-frequency acquisition mode to high-frequency acquisition mode when the monitoring system infers that the anchor bolt load change exceeds the load change threshold.

5. The machine vision-based deformation detection method for anti-buoyancy anchor bolts as described in claim 1, characterized in that, In step S2, the internal and external parameters of the industrial camera need to be pre-calibrated. The specific steps are as follows: A stereo calibration plate with known precise three-dimensional dimensions is installed on a fixed structure near the anchor bolt; the industrial camera is controlled to capture images containing the stereo calibration plate from multiple different angles to form a calibration image set; The image coordinates of feature points on the stereo calibration plate are detected from the calibration image set; combined with the known world coordinates of the feature points on the stereo calibration plate, the intrinsic parameter matrix, distortion coefficient, and extrinsic parameter matrix of the industrial camera relative to the coordinate system of the stereo calibration plate are solved by minimizing the reprojection error; and a fixed transformation relationship is established between the coordinate system of the industrial camera and the local three-dimensional coordinate system of the anchor bolt through coordinate system transformation.

6. The machine vision-based deformation detection method for anti-buoyancy anchor bolts as described in claim 1, characterized in that, In step S3, the working principle of the improved template matching algorithm includes: S31. In the reference image, take the center of each high-contrast artificial marker point as the origin and extract a standard-sized image block as the initial template. S32. In subsequent images to be detected, a search area larger than the initial template is defined, centered on the theoretically predicted position of the high-contrast artificial marker. S33. Within the search area, normalized cross-correlation is used to calculate the similarity between the initial template and each sub-region of the image to be detected, and a similarity distribution map is obtained. S34. In the similarity distribution map, locate the similarity peak point and preliminarily determine the candidate positions of high-contrast artificial marker points; S35. Extract the image patch centered on the candidate position, and input it together with the initial template into a lightweight convolutional neural network for feature re-extraction to obtain a deep feature descriptor; S36. Calculate the cosine similarity between the deep feature descriptors. When the cosine similarity is higher than a set threshold, confirm the candidate position as the pixel position of the high-contrast artificial marker.

7. The machine vision-based deformation detection method for anti-buoyancy anchor bolts as described in claim 5, characterized in that, In step S5, converting the two-dimensional pixel displacement vector to the local three-dimensional space of the anchor surface includes: S51. Using the camera intrinsic parameter matrix calibrated by the industrial camera in step 2, eliminate the influence of image distortion on pixel coordinates and obtain the normalized camera coordinates of each high-contrast artificial marker point. S52. Based on the extrinsic parameter matrix of the industrial camera relative to the coordinate system of the anchor bolt observation section, transform the normalized camera coordinates to the local three-dimensional coordinate system of the anchor bolt surface; S53. In the local three-dimensional coordinate system, the three-dimensional coordinates of each high-contrast artificial marker point in the reference image are known, and the three-dimensional coordinates of the corresponding point in the current frame image are obtained through the aforementioned coordinate transformation. S54. Calculate the coordinate difference between the same high-contrast artificial marker point in the current frame and the reference image in the local three-dimensional coordinate system. The coordinate difference is the three-dimensional spatial displacement of the high-contrast artificial marker point.

8. The machine vision-based deformation detection method for anti-buoyancy anchor bolts as described in claim 1, characterized in that, In step S7, after reversing the axial strain distribution and lateral deflection curve of the anchor rod, a deformation state assessment step is also included: extracting the axial strain distribution curve along the entire length of the anchor rod from the final inversion result of the parameterized elastic foundation beam flexural deformation analysis model; extracting the lateral deflection curve along the entire length of the anchor rod from the final inversion result; calculating the first derivative of the lateral deflection curve to obtain the angle change curve of the anchor rod; calculating the second derivative of the lateral deflection curve, and deriving the bending moment distribution curve of the anchor rod section by combining the elastic modulus information of the anchor rod.

Images