Bridge template deformation intelligent monitoring method fused with machine vision

Abstract

The invention discloses a bridge template deformation intelligent monitoring method fused with machine vision, and belongs to the crossing field of computer image processing and civil engineering monitoring technologies, and the method comprises the steps: generating a theoretical edge feature map according to an obtained three-dimensional model and calibration parameters, carrying out the distance transformation of the theoretical edge feature map, and generating a two-dimensional distance potential energy field; collecting a video stream, extracting a current video frame, and generating an actually measured edge feature map; a reverse projection objective function is constructed and solved based on the actually measured edge feature map and the two-dimensional distance potential energy field, a total displacement field is generated, orthogonal decomposition is executed, and a rigid body motion component, an accumulated plastic deformation component and a local elastic deformation component are decoupled; and generating a deformation visualization result and a safety early warning signal based on the local elastic deformation component and the accumulated plastic deformation component. A total displacement field is solved by adopting reverse projection based on a three-dimensional model, and an orthogonal decomposition scheme is carried out on the total displacement field, so that rigid body motion and real structure deformation can be separated, and full-field deformation monitoring of the bridge template is realized.

Description

Technical Field

[0001] This invention relates to the field of intersection between computer image processing and civil engineering monitoring technology, and in particular to an intelligent monitoring method for bridge formwork deformation that integrates machine vision. Background Technology

[0002] Bridge formwork is a critical temporary support system in the construction of bridge superstructures. Its stability and shape under heavy loads such as concrete pouring directly affect the final bridge's alignment accuracy and structural safety. Therefore, real-time and accurate monitoring of bridge formwork deformation during construction is essential. Utilizing machine vision technology for non-contact monitoring has become an important development direction in this field.

[0003] Among related technologies, Chinese invention patent CN120160585B discloses a method for monitoring the deformation of pumped storage dams based on BeiDou positioning. The method includes: deploying a dual-frequency BeiDou receiver array and environmental sensors; fusing BeiDou observation data, reservoir water pressure gradient, foundation vibration spectrum and three-dimensional geological structure data; constructing a robust solution model constrained by a geological model; outputting the confidence interval of deformation parameters and suppressing outliers through robust estimation; and using a dynamic time warping algorithm to match historical operating condition databases, combined with a closed-loop feedback mechanism to dynamically calibrate early warning thresholds and reverse-optimize the monitoring network.

[0004] However, the aforementioned monitoring schemes based on satellite positioning or contact sensors can typically only acquire displacement data from sparse, discrete points, making it difficult to precisely capture localized bulges or continuous curvature changes in bridge formwork within stress concentration areas. Furthermore, the large size of the equipment makes high-density deployment in space-constrained scaffolding systems difficult. While existing machine vision methods possess the potential for full-field monitoring, most rely on attaching high-contrast artificial markers (targets) to the measured surface or using the rich texture of the object's surface for optical flow tracing. For steel formwork with smooth surfaces, limited texture features, and subjected to strong vibrations during concrete pouring, artificial markers are easily obscured by mud splashes or construction scratches. Conventional vision algorithms also struggle to effectively distinguish between camera shake (rigid body motion) and the actual deformation of the formwork (elastic deformation), resulting in low signal-to-noise ratios in the monitoring data and failing to meet the requirements for high-precision, real-time safety warnings at construction sites. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an intelligent monitoring method for bridge formwork deformation that integrates machine vision. It employs a technical solution based on the inverse projection of a 3D model to solve the total displacement field and perform orthogonal decomposition, which can effectively separate rigid body motion from the actual structural deformation, thereby achieving high-precision full-field deformation monitoring of bridge formwork.

[0006] The above objectives can be achieved through the following approach: A method for intelligent monitoring of bridge formwork deformation integrating machine vision includes: acquiring a 3D model describing the geometry of the bridge formwork and calibration parameters describing the camera imaging relationship; generating a theoretical edge feature map based on the 3D model and calibration parameters; performing a distance transformation on the theoretical edge feature map to generate a two-dimensional distance potential energy field; extracting the current video frame from the video stream captured by the on-site camera and extracting a measured edge feature map from the current video frame; constructing and solving an inverse projection objective function based on the measured edge feature map and the two-dimensional distance potential energy field to generate a total displacement field; performing orthogonal decomposition on the total displacement field to decouple rigid body motion components, cumulative plastic deformation components, and local elastic deformation components; and generating deformation visualization results and safety warning signals based on the local elastic deformation components and cumulative plastic deformation components.

[0007] Optionally, generating the theoretical edge feature map includes: acquiring an initial deformation field to simulate the minute deformation of a bridge template under initial stress; applying the initial deformation field to the three-dimensional model to generate a set of candidate three-dimensional models; rendering each candidate three-dimensional model in the candidate three-dimensional model set according to the calibration parameters to generate a candidate benchmark feature map set; extracting the first frame image of the video stream and processing the first frame image to generate a first frame edge feature map; calculating the image quality confidence weight of each region in the first frame edge feature map, and generating a comprehensive matching degree based on the matching degree between the image quality confidence weight and the first frame edge feature map and the candidate benchmark feature map set; and selecting the feature map with the highest matching degree from the candidate benchmark feature map set as the theoretical edge feature map based on the comprehensive matching degree.

[0008] Optionally, generating the total displacement field includes: determining the initial state of the three-dimensional model based on the theoretical edge feature map and initializing vertex displacements; generating a calculated edge feature map by forward projection of the current vertex displacements and the calibration parameters; comparing the calculated edge feature map with the measured edge feature map to generate a residual map; analyzing the spatial distribution pattern of the residual map and dynamically adjusting the regional weights of the two-dimensional distance potential energy field based on the spatial distribution pattern to generate an adaptive distance potential energy field; and iteratively optimizing vertex displacements based on the smoothness physical constraints of the adaptive distance potential energy field and adjacent vertex displacements to generate the total displacement field.

[0009] Optionally, generating the adaptive distance potential field includes: identifying the distribution pattern of pixels in the residual image and classifying it into a random noise pattern, a directional stripe pattern, or a structural continuous pattern; obtaining the image regions corresponding to the random noise pattern and the directional stripe pattern, and generating a weighting factor to reduce the influence of the image regions in the two-dimensional distance potential field; and applying the weighting factor to the two-dimensional distance potential field to generate the adaptive distance potential field.

[0010] Optionally, the method further includes: obtaining the convergence speed of the iterative optimization vertex displacement process, and combining the spatial distribution pattern of the residual map with the image quality confidence weight to comprehensively calculate and generate the optimization result confidence of the current frame; and associating the optimization result confidence with the total displacement field to generate displacement data with confidence labels.

[0011] Optionally, performing orthogonal decomposition on the total displacement field includes: acquiring multiple total displacement fields generated from continuous video frames to construct displacement time series for each vertex of the three-dimensional model; performing time-frequency analysis on the displacement time series to decompose it into high-frequency displacement components, mid-to-low-frequency displacement components, and cumulative plastic deformation components; obtaining the time points of vibration events generated by external construction equipment, and filtering out instantaneous interference components by associating the vibration event time points with the occurrence time of the high-frequency displacement components; analyzing the spatial distribution consistency of the mid-to-low-frequency displacement components to separate rigid displacement components and local elastic deformation components.

[0012] Optionally, the generation of deformation visualization results and safety warning signals includes: vector synthesis based on the local elastic deformation components and the cumulative plastic deformation components to generate a net deformation field; obtaining a safety deformation threshold for defining the deformation hazard level; comparing the magnitude of the net deformation field with the safety deformation threshold to trigger a safety warning signal, and rendering the net deformation field onto the surface of the three-dimensional model to generate the deformation visualization results.

[0013] Optionally, the method further includes: evaluating the overall confidence level of the net deformation field; obtaining a confidence threshold for judging whether the result is reliable; when the overall confidence level is higher than the confidence threshold, inversely fusing the cumulative plastic deformation component into the three-dimensional model to generate an updated three-dimensional reference model, and using the updated three-dimensional reference model to perform subsequent monitoring procedures.

[0014] Optionally, generating the updated 3D reference model includes: obtaining a scaling factor for smoothing the model update process; attenuating the cumulative plastic deformation component according to the scaling factor to generate attenuated nodal displacements; and applying the attenuated nodal displacements to the corresponding vertices of the 3D model to update the geometric data of the 3D model, thereby generating the updated 3D reference model.

[0015] Based on the same inventive concept, this invention also provides an intelligent monitoring system for bridge formwork deformation that integrates machine vision, the system comprising: The benchmark construction module is used to obtain a 3D model to describe the geometry of the bridge template and calibration parameters to describe the camera imaging relationship. Based on the 3D model and calibration parameters, a theoretical edge feature map is generated. A distance transformation is performed on the theoretical edge feature map to generate a two-dimensional distance potential energy field. The feature extraction module is used to extract the current video frame from the video stream captured by the on-site camera, and to extract the measured edge feature map from the current video frame; The deformation calculation module is used to construct and solve the inverse projection objective function based on the measured edge feature map and the two-dimensional distance potential energy field, and generate the total displacement field. The decoupling analysis module is used to perform orthogonal decomposition on the total displacement field to decouple the rigid body motion component, the cumulative plastic deformation component and the local elastic deformation component. The output warning module is used to generate deformation visualization results and safety warning signals based on the local elastic deformation component and the cumulative plastic deformation component.

[0016] Compared with the prior art, the present invention has the following advantages: 1. By constructing an initial deformation candidate set and matching it with the first frame image, the difference between the ideal model and the initial state on site is effectively compensated. Furthermore, an adaptive weighting mechanism based on residual pattern analysis is introduced to dynamically suppress visual interference such as illumination changes and environmental noise, thereby improving the accuracy and robustness of the deformation solution results.

[0017] 2. By performing time-frequency analysis and spatial consistency analysis on the total displacement field obtained from continuous monitoring, the original displacement data with mixed physical meanings was successfully orthogonally decomposed into independent components such as rigid body motion, cumulative plastic deformation, and local elastic deformation. This enabled the monitoring system to accurately distinguish between false displacements caused by camera jitter and true deformations that truly reflect the structural health, providing a clean and reliable data foundation for safety assessment.

[0018] 3. An adaptive model update closed-loop feedback mechanism was established. After confirming the occurrence of high-confidence cumulative plastic deformation, the permanent deformation amount can be back-integrated into the three-dimensional reference model, thereby dynamically adjusting the monitoring benchmark. This effectively avoids continuous false alarms caused by permanent changes in structural morphology and enhances the applicability and reliability during long-term service.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an intelligent monitoring method for bridge template deformation that integrates machine vision, according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of an intelligent monitoring method for bridge template deformation that integrates machine vision, according to an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of an intelligent monitoring method for bridge template deformation that integrates machine vision, according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of an intelligent monitoring system for bridge template deformation that integrates machine vision, according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0026] Reference Figure 1 One embodiment of the present invention proposes an intelligent monitoring method for bridge template deformation that integrates machine vision. It adopts a technical solution of solving the total displacement field by inverse projection based on a three-dimensional model and performing orthogonal decomposition on it, which can effectively separate rigid body motion from the actual structural deformation and realize high-precision full-field deformation monitoring of bridge template.

[0027] The method described in this embodiment specifically includes: A three-dimensional model describing the geometry of the bridge template and calibration parameters describing the camera imaging relationship are obtained. A theoretical edge feature map is generated based on the three-dimensional model and calibration parameters. A distance transformation is performed on the theoretical edge feature map to generate a two-dimensional distance potential energy field. Extract the current video frame from the video stream captured by the on-site camera, and extract the measured edge feature map from the current video frame; Based on the measured edge feature map and the two-dimensional distance potential field, an inverse projection objective function is constructed and solved to generate the total displacement field; Perform orthogonal decomposition on the total displacement field to decouple the rigid body motion component, the cumulative plastic deformation component, and the local elastic deformation component; Based on the local elastic deformation component and the cumulative plastic deformation component, deformation visualization results and safety warning signals are generated.

[0028] Specifically, using a priori 3D model of the bridge template and camera calibration parameters, a theoretically idealized edge image is generated in the forward direction. This image is then used as a benchmark to construct a two-dimensional distance potential energy field, which provides target guidance for subsequent matching. Subsequently, by matching the actual edge images collected on-site with this potential energy field, an inverse projection objective function is constructed and solved, thereby calculating the total displacement field describing the overall displacement of the template. To accurately analyze the structural state, this method further performs orthogonal decomposition on the total displacement field, separating rigid body motion, cumulative plastic deformation, and local elastic deformation, which have different physical meanings. Finally, the deformation components directly related to structural safety are used to generate monitoring results. This achieves full-field, non-contact intelligent monitoring of bridge template deformation, overcoming the limitation of traditional single-point monitoring methods that cannot comprehensively reflect the structural state. By orthogonally decomposing the calculated total displacement field, interference from rigid body motion caused by non-structural factors such as camera displacement can be effectively filtered out, thereby accurately identifying the local elastic deformation and cumulative plastic deformation that truly reflect the stress state of the template. This method improves the reliability of monitoring results and the accuracy of early warning signals, avoiding false alarms caused by environmental interference. By generating intuitive deformation visualization results and timely safety early warning signals, it can provide clear and reliable decision-making basis for on-site construction safety management, thereby effectively ensuring construction safety.

[0029] Optionally, the generation of the theoretical edge feature map includes: Obtain the initial deformation field for simulating the minute deformation of bridge templates under initial stress; The initial deformation field is applied to the three-dimensional model to generate a set of candidate three-dimensional models; Each candidate 3D model in the candidate 3D model set is rendered according to the calibration parameters to generate a candidate baseline feature map set; Extract the first frame image of the video stream and process the first frame image to generate the first frame edge feature map; Calculate the image quality confidence weights for each region of the first frame edge feature map, and generate a comprehensive matching score based on the matching degree between the image quality confidence weights and the first frame edge feature map and the candidate baseline feature map set. Based on the overall matching degree, the feature map with the highest matching degree is selected from the candidate benchmark feature map set as the theoretical edge feature map.

[0030] Specifically, an initial deformation field is acquired to simulate initial minute deformations. This initial deformation field is not a single deformation mode, but rather consists of a set of pre-calculated structural principal vibration modes or typical stress modes based on finite element analysis, such as first-order vertical bending and first-order torsion. The deformation amplitude is limited to a small range consistent with engineering realities, for example, a maximum displacement of no more than 5 millimeters. Different deformation modes in this initial deformation field are linearly superimposed with different weighting coefficients and applied to the vertices of the original 3D model, generating a set of candidate 3D models containing multiple possible initial forms. Based on pre-calibrated camera calibration parameters, each model in the candidate 3D model set is virtually rendered to generate a corresponding 2D projected image. Edge extraction operations are performed on these images, for example, using the Canny edge detection operator, to construct a candidate baseline feature map set containing multiple candidate edge images. Simultaneously, the first frame image is captured from the video stream acquired by the field camera, and the same edge extraction algorithm is used to generate the first frame edge feature map representing the actual physical edges at the start of monitoring. The first frame image is divided into a grid, and the image gradient variance or local contrast within each grid region is calculated to generate image quality confidence weights reflecting imaging quality. Regions with high weight values ​​indicate clear images and reliable features, thus holding greater weight in subsequent matching. The optimal benchmark is selected from the candidate set by calculating the overall matching degree. The formula for calculating the overall matching degree S_k is as follows: ; in, represents the overall matching degree of the k-th candidate baseline feature map. p represents each pixel in the image. It represents the image quality confidence weight for the corresponding pixel, obtained from image quality analysis. This is the edge identifier of pixel p in the edge feature map of the first frame. If p is an edge point, it is 1; otherwise, it is 0. Let G be the edge identifier of pixel p in the k-th candidate benchmark feature map. G is a pixel-level similarity function, for example: ; Used to penalize mismatched pixels. Iterate through all candidate baseline feature maps, calculate their comprehensive matching score, and select the candidate baseline feature map with the highest comprehensive matching score as the final theoretical edge feature map.

[0031] For example, a steel box girder bridge formwork with a span of 30 meters is selected as the monitoring object. First, an initial deformation field is constructed, and two principal vibration modes from the finite element analysis output are set: first-order vertical bending (mode 1) and first-order torsion (mode 2). Based on engineering experience and the elastic range of steel structures, the maximum displacement amplitude is limited to 5 mm. Three candidate models are generated: Model A weight coefficient... Apply mode 1 (maximum mid-span deflection 5mm) and model B weighting coefficients. Apply mode 2, model C is Taking model A as an example, its mid-span vertex The initial coordinates are meters, superimposed displacement vector After meters, the vertex becomes Meters. Using the camera intrinsic parameter matrix. and external references Model A is rendered to generate a projection map with a resolution of 1920x1080. Edges are extracted using the Canny operator with a high threshold of 100 and a low threshold of 50 to obtain candidate feature maps. Simultaneously, the first frame of the live video is acquired, and the true edge feature map R is extracted. The image is then divided into... Calculate the gradient variance within the pixel grid. ,like , The empirical threshold represents a clear texture, thus determining the confidence weight for that region. ,otherwise Regarding the formula Verification: In a certain The weight of the center pixel p within a local region First frame edge map Candidate graph A at this point Similarity Then the contribution score for that point is If candidate graph B is at this point Similarity The contribution score is 0. After traversing the entire graph and accumulating the scores, if... , , Select The corresponding model A serves as the theoretical edge feature map. This process ensures that the baseline model best reflects physical reality through quantized matching scores.

[0032] Optionally, generating the total displacement field includes: The initial state of the 3D model is determined based on the theoretical edge feature map, and the vertex displacements are initialized. An edge feature map is generated by projecting the current vertex displacement onto the calibration parameters. The calculated edge feature map is compared with the measured edge feature map to generate a residual map; Analyze the spatial distribution pattern of the residual map, and dynamically adjust the regional weights of the two-dimensional distance potential field according to the spatial distribution pattern to generate an adaptive distance potential field; Based on the physical constraints of the adaptive distance potential energy field and the smoothness of the adjacent vertex displacements, the vertex displacements are iteratively optimized to generate the total displacement field.

[0033] Specifically, based on the theoretical edge feature map, the corresponding 3D model state is set as the initial state for deformation calculation, and the displacement vectors of all vertices are initialized to zero. An iterative optimization loop is then initiated for the current video frame. In each iteration, forward projection is first performed: the current 3D vertex displacements are applied to the initial 3D model, and the deformed model is projected onto the 2D image plane using camera calibration parameters. Then, edge feature maps are generated through edge extraction. This calculated edge feature map is compared pixel-level with the measured edge feature map extracted from the current video frame to generate a residual map. This map visually shows the mismatch between the current model prediction and the actual observation. The spatial distribution patterns of the residual map are analyzed, such as identifying random noise caused by changes in illumination or water reflection, or directional stripes caused by slight camera vibrations. Based on these patterns, a region weight map is generated, assigning lower weights (e.g., 0.1 to 0.3) to residual regions with low confidence, and higher weights (e.g., 0.8 to 1.0) to regions exhibiting structural continuous differences that may indicate true deformation. This region weight map is used to dynamically adjust the adaptive distance potential field. A basic potential energy field is obtained by performing a distance transformation on the measured edge feature map, where the value of each pixel represents its distance to the nearest ground truth edge. This basic potential energy field is then multiplied pixel-by-pixel with the aforementioned region weight map to generate an adaptive distance potential energy field. Based on the adaptive distance potential energy field and the physical constraints ensuring continuous and smooth deformation, an objective function is constructed and optimized to solve for the final total displacement field. The mathematical expression of this objective function is: ; Here, D represents the total displacement field of the entire model, which consists of the displacement vectors of each vertex. Composition. The first item is the data item. It is the initial vertex of the model After displacement The two-dimensional coordinates of the projection onto the image are then displayed. The first term is the value of the adaptive distance potential field at that point, designed to drive the model projection closer to the measured edge. The second term is a smoothness physical constraint, where N represents the set of all adjacent vertices in the 3D model, and λ is a regularization parameter used to balance data matching and model smoothness; its engineering value is typically between 0.1 and 1.0. This term penalizes excessive relative displacement between adjacent vertices, ensuring that the calculated deformation is physically reasonable. A nonlinear optimization algorithm, such as the Gauss-Newton method, is used to iteratively solve for the displacement field D that minimizes the objective function E. The iteration terminates when the change in the objective function value is less than a preset convergence threshold, and the resulting displacement field D is the total displacement field of the current frame.

[0034] For example, we proceed to iterative optimization. The current iteration step is 1, and the initial displacement is... Regarding the formula Detailed derivation and verification are performed. Parameter settings: Regularization parameter. The parameters are determined empirically, balancing data fit and smoothness. Data item calculation: Consider a vertex in the model. Current estimated displacement Vertex after deformation The point is projected onto the image coordinates using the projection function P. Query the adaptive distance potential field The coordinates If the Euclidean distance to the nearest ground truth edge is 2 pixels, and the weight of this region is 1.0, then... Smoothing term calculation: Let... Adjacent vertices Its current displacement Then the square of the displacement difference norm Total value of the objective function: For only these two points, then... The Jacobian matrix is ​​calculated using the Gauss-Newton method. In the 5th iteration, the objective function value is... The difference (With a preset convergence threshold), the iteration stops, and the current D is output as the total displacement field. This process ensures that the displacement calculation conforms to both visual observation and structural continuity.

[0035] Optionally, generating the adaptive distance potential field includes: Identify the distribution pattern of pixels in the residual image and classify it into random noise pattern, directional stripe pattern, or structural continuous pattern; Obtain the image regions corresponding to the random noise pattern and the directional stripe pattern, and generate a weighting factor to reduce the influence of the image regions in the two-dimensional distance potential field. The weighting factor is applied to the two-dimensional distance potential field to generate an adaptive distance potential field.

[0036] Specifically, in each iteration of the deformation calculation, spatial distribution pattern analysis is performed on the resulting residual map. The first step of this analysis process is to identify and classify the distribution patterns of non-zero pixels in the residual map. Using image processing algorithms, these pixel patterns are categorized into three types: random noise patterns, characterized by isolated, small-area pixel clusters, usually caused by sensor noise or slight light flicker; directional stripe patterns, characterized by elongated pixel clusters with obvious directionality, typically originating from high-frequency mechanical vibration or camera scan line interference; and structural continuity patterns, i.e., large continuous pixel regions related to the bridge template outline, which are usually valid signals of actual structural deformation. The algorithm automatically delineates the image regions corresponding to the random noise pattern and the directional stripe pattern. For example, connected component analysis can be applied to identify connected components with an area smaller than a specific threshold, such as 10 pixels, as random noise regions; simultaneously, Fourier transform or histogram of directional gradients can be used to detect the direction of energy concentration, and pixel regions distributed along this direction are identified as directional stripe regions. For image regions identified as unreliable signals, weighting factors are generated to reduce their impact in subsequent calculations. These weighting factors are values ​​between 0 and 1; lower values ​​(e.g., 0.2) are set for noisy regions with high confidence, while a weighting factor of 1.0 is maintained for regions with structurally continuous patterns. These spatially varied weighting factors are then multiplied pixel-by-pixel and applied to an initial two-dimensional range potential field generated based on the measured edge feature map of the current frame, thus generating the final adaptive range potential field. This process can be expressed by the following formula: ; Where p represents the pixel coordinates in the image. It is the value of the generated adaptive distance potential field at point p. It is the original potential field value generated by distance transformation based on the measured edge feature map of the current video frame. The weighting factor applied to point p is obtained through the residual plot pattern analysis described above.

[0037] For example, the measured edge feature map extracted in the current frame contains noise due to water surface reflection. First, the original distance potential field is generated. For image coordinates The distance from the nearest edge pixel is d=10. Pattern analysis and weight generation: Connectivity analysis is performed on the residual map. It is found that the connected area of ​​the non-zero pixel region where coordinate p is located is 5 pixels. The random noise threshold is set to 10 pixels, that is, an area less than 10 pixels is considered noise. This region is determined to be a random noise pattern. Weighting factors for random noise patterns. If a region is a continuous long strip, meaning the length of the line detected by Hough transform is greater than 50 pixels, it is determined to be a structural pattern, and the weight is adjusted accordingly. Formula calculation: According to the formula For the noise point p mentioned above, its adaptive potential energy value is In contrast, without weighting, the potential energy at this point is 10, which would generate a huge pull in the optimization function, incorrectly drawing the model toward the noise. After weighting, the attraction at this point is reduced by 80%, thus achieving anti-interference. This step is based on the fact that small, isolated noise points usually do not have structural physical significance and their contribution should be reduced in the optimization.

[0038] Optionally, the method further includes: The convergence speed of the iterative optimization vertex displacement process is obtained, and the confidence score of the optimization result of the current frame is calculated by combining the spatial distribution pattern of the residual map and the image quality confidence weight. The confidence level of the optimization result is correlated with the total displacement field to generate displacement data with confidence level labels.

[0039] Specifically, after completing the iterative optimization of vertex displacement, the convergence speed of this process is immediately obtained, usually measured by the number of iterations required to reach convergence. A smaller number of iterations, such as less than 20, generally indicates that the model matches the measured edges well and quickly, suggesting high confidence; conversely, a larger number of iterations may mean that the matching process is ambiguous or trapped in a local optimum. Simultaneously, the spatial distribution pattern of the final residual map after optimization is analyzed, and a residual quality metric is calculated. The residual quality metric aims to quantify whether the residuals tend towards random noise or a structurally continuous pattern. For example, the average area of ​​the residual region or its energy distribution concentration on the Fourier spectrum can be calculated. An ideal residual map should resemble low-energy random scattering points, corresponding to a higher residual quality metric value. Image quality confidence weights related to the current matching region are aggregated to obtain a comprehensive image quality index. The confidence score of the optimization result for the current frame is generated using a comprehensive calculation formula: ; in, is the confidence score of the final output optimization result, with a value range of 0 to 1. N, M, and W are the number of convergence iterations, the residual quality measure, and the overall image quality index, respectively. , , These are three normalization functions, which map their respective input physical quantities to a unified confidence interval of 0 to 1. For example, the inverse proportional function decreases as N increases. , , These are preset weighting coefficients, the sum of which is 1. Their specific values, such as 0.3, 0.5, and 0.2, are determined based on historical experience, reflecting different emphases on convergence speed, residual quality, and image quality when evaluating the reliability of the results. After calculating the confidence level of the optimization result, it is used as a label and correlated with the total displacement field calculated in the current frame. This ultimately generates a set of displacement data with confidence labels, providing crucial information for subsequent deformation analysis and data selection.

[0040] For example, to quantify the reliability of the current frame's solution, all normalization functions need to be specified. Parameter settings: weight coefficients. Function definition and calculation: Convergence rate term Let the optimal number of iterations be... If the current frame iteration count ,but This function is based on the principle that the more iterations, the greater the uncertainty of the solution, showing an inverse relationship. Residual quality term. M is defined as the average area of ​​connected components of non-zero pixels in the residual graph, i.e., the number of pixels. The residual is expected to be fragmented, i.e., noise. A normalization function is defined. If unmatched structural fringes appear in the residual plot, resulting in an average connected area M=20, then... If the residuals are only scattered points, and M=2, then Here, we use M=20 for calculation. Image quality item. W represents the average confidence weight of the pixels participating in the matching. If the current lighting is good, the average weight W = 0.8, which is directly defined. Substitute into the formula The result of 0.494 is lower than the usual confidence threshold of 0.7, so the system will mark the frame of data as low confidence, indicating that there may be a measurement error.

[0041] Optionally, performing orthogonal decomposition on the total displacement field includes: Multiple total displacement fields generated from consecutive video frames are acquired to construct displacement time series for each vertex of the three-dimensional model; The displacement time series was subjected to time-frequency analysis and decomposed into high-frequency displacement components, mid-to-low frequency displacement components, and cumulative plastic deformation components. The time points of vibration events generated by external construction equipment are obtained, and transient interference components are filtered out by correlating the vibration event time points with the occurrence time of high-frequency displacement components. The spatial distribution consistency of low- and medium-frequency displacement components was analyzed, and rigid displacement components and local elastic deformation components were separated.

[0042] Specifically, multiple total displacement fields are calculated from continuous video frames, for example, at a frequency of 1 Hz for one hour, to construct a displacement time series for each vertex in the 3D model. Time-frequency analysis is performed on the displacement time series of each vertex, using wavelet transform or empirical mode decomposition to decompose it into a superposition of different frequency components. This decomposition process can be expressed as: ; in, It is the total displacement at time t. It is a high-frequency displacement component, whose frequency is usually greater than 1 Hz, mainly reflecting the instantaneous shaking caused by environmental vibration or construction machinery. It is a low-to-medium frequency displacement component, with a frequency range between 0.001 Hz and 1 Hz, corresponding to the structural elastic response caused by concrete pouring loading, temperature changes, etc. The cumulative plastic deformation component, manifested as a long-term trend or baseline drift in the time series, represents the permanent and irreversible deformation of the structure. The time points of vibration events generated by external construction equipment, such as vibrators and transport vehicles, are acquired. By correlating these time points with the times of energy peaks in the high-frequency displacement components, transient interference components caused by known external sources are identified and filtered out. After high-frequency filtering, the spatial distribution consistency of the mid- and low-frequency displacement components across the entire bridge formwork is analyzed. Through global rigid body transformations—translation and rotation—the mid- and low-frequency displacements of all vertices are fitted. The portion perfectly explained by this rigid body transformation model is separated into rigid displacement components; while the fitted residual displacement, i.e., the relative displacement of each vertex relative to this global motion, is defined as a local elastic deformation component. Through this series of steps, the original total displacement field is successfully decoupled into physically meaningful rigid body motion components, cumulative plastic deformation components, and local elastic deformation components.

[0043] For example, displacement data was collected for one hour at a sampling rate of 30Hz, resulting in 108,000 data points per vertex. The mid-span vertex of the model was selected. Its total displacement time series A 5-level decomposition was performed using the db4 wavelet. The frequency was then... Detail coefficients reconstructed as Suppose that at t=500s, a high-frequency oscillation with an amplitude of 2mm is detected. Corresponding to the construction log, it is found that a vibratory roller was operating at this time. This component is confirmed as interference and is filtered out. For the remaining low-frequency signal... Perform linear trend fitting If the calculated slope a = 0.001 mm / s and intercept b = 0, then at t = 3600 s, the cumulative plastic deformation is... The remaining fluctuation components This includes rigid body motion and elastic deformation. Three reference points on the model are selected, assumed to be at the bridge piers, where the theoretical displacement should be 0. The average displacement vector is calculated as the overall rigid body translation $T = (0.5, 0.2, 0)$ mm. For the vertices... Its low-frequency displacement at time t is (5.5, 0.2, 0) mm. Subtracting the rigid body translation T, we obtain the local elastic deformation. .

[0044] Optionally, the generation of deformation visualization results and safety warning signals includes: A net deformation field is generated by vector synthesis of the local elastic deformation component and the cumulative plastic deformation component. Obtain the safe deformation threshold used to define the deformation hazard level; The net deformation field is compared with the safe deformation threshold to trigger a safety warning signal, and the net deformation field is rendered onto the surface of the three-dimensional model to generate the deformation visualization result.

[0045] Specifically, based on the local elastic deformation components and the cumulative plastic deformation components, vector synthesis is performed on each vertex of the 3D model. This synthesis operation involves directly adding the two displacement vectors of the corresponding vertex to generate a net deformation field that comprehensively reflects the current true structural state. This process can be expressed by the formula: ; in, It is the net deformation vector of vertex i in the model. and These are the local elastic deformation component and the cumulative plastic deformation component vectors of the vertex, respectively. The safety deformation thresholds, preset according to bridge design specifications and construction plans, are obtained from the configuration file. These thresholds are typically graded, for example, including a first-level warning threshold of 10 mm, a second-level alarm threshold of 15 mm, and a limit threshold representing a dangerous structural state of 20 mm. The magnitude of the displacement vector of each vertex in the net deformation field, i.e., its Euclidean norm, is calculated in real time, and the maximum deformation value on the entire model is found. This maximum deformation value is continuously compared with the graded safety deformation thresholds. Once the value exceeds a certain level of threshold, the corresponding safety warning signal is immediately triggered. This signal can be manifested as sending an alarm SMS to the manager's mobile phone, displaying a highlighted warning box on the monitoring center software interface, or activating the on-site audible and visual alarm. The calculated net deformation field is rendered onto the surface of the 3D model to generate deformation visualization results. This rendering process is achieved by applying a color mapping table, mapping the magnitude of the deformation to different colors. For example, areas with deformation less than 5 mm are rendered in green, areas between 5 and 15 mm are rendered in yellow, and areas exceeding 15 mm are highlighted in striking red, allowing managers to clearly understand the size, distribution, and degree of danger of the bridge formwork deformation at a glance.

[0046] For example, for vertices A comprehensive evaluation is conducted. If at the current moment, the vertex... Local elastic deformation vector That is, downward elastic deformation, cumulative plastic deformation vector That is, permanent downward scratching. According to the formula... Calculate the net deformation modulus. The system configuration file sets the threshold for Level 1 warning (yellow) to 10mm and the threshold for Level 2 alarm (red) to 15mm. Since 10mm < 12mm < 15mm, the system is currently in Level 1 warning mode. The rendering engine is invoked to look up the color mapping table: [0-5mm: green, 5-10mm: blue, 10-15mm: yellow, >15mm: red]. Vertex The signal is rendered in yellow. Simultaneously, a signal with the code "WARN_L1" is sent to the control center, triggering a yellow warning box to pop up on the monitoring screen, displaying "Structural deformation detected: 12mm, exceeding the first-level threshold." This logic ensures that monitoring data can be instantly transformed into intuitive management signals.

[0047] Optionally, the method further includes: Assess the overall reliability of the net deformation field; A confidence threshold is obtained to determine whether the result is reliable. When the overall confidence level is higher than the confidence threshold, the cumulative plastic deformation component is back-fused into the three-dimensional model to generate an updated three-dimensional reference model. The updated three-dimensional reference model is then used to perform subsequent monitoring procedures.

[0048] Specifically, after generating the net deformation field, its overall reliability is evaluated. This overall reliability is obtained by aggregating the confidence scores of each frame of optimization results within the time window for calculating the cumulative plastic deformation component. For example, a weighted average or minimum value of the built-in confidence scores for the time period is used to ensure that the solution quality remains at a high level throughout the deformation process. A confidence threshold for judging the reliability of the results is obtained from the configuration. This threshold is a high standard value set based on engineering experience, such as 0.85. The calculated overall reliability is compared with this confidence threshold. If and only if the overall reliability is higher than the confidence threshold, the currently calculated cumulative plastic deformation component is determined to be real and reliable, and the model update process is triggered. In this process, the verified cumulative plastic deformation component is back-fused into the current 3D reference model. This fusion process is not a one-time abrupt change, but rather the displacement vector is used as a correction amount and applied to the geometric coordinates of the corresponding vertices of the model, thereby generating an updated 3D reference model with a geometry that is closer to the current physical reality. Once the update is complete, this updated 3D reference model will replace the original reference model and be used to perform all subsequent monitoring processes, such as the generation of theoretical edge feature maps and displacement field calculations, all based on this new model.

[0049] For example, assume the evaluation time window for model updates is set to 10 minutes, or 600 frames. During this period, the calculated cumulative plastic deformation... The results are trending towards stability. The confidence level of the optimization results across these 600 frames is being collected. Calculate its weighted average. Assuming most frames are sharp and converge quickly, Retrieve the confidence threshold from the configuration file. The comparison revealed that 0.88 > 0.85, satisfying the high confidence condition. It was determined that the currently calculated plastic deformation is a real physical phenomenon, not an algorithm error. The current cumulative plastic deformation field data was locked, and a command was sent.

[0050] Optionally, generating the updated 3D reference model includes: Obtain the scaling factor used to smooth the model update process; The cumulative plastic deformation component is attenuated according to a proportionality factor to generate attenuated nodal displacements. The attenuated node displacements are applied to the corresponding vertices of the 3D model to update the geometric data of the 3D model, generating an updated 3D reference model.

[0051] Specifically, after the high-confidence trigger condition is met, a scaling factor used to smooth the model update process is obtained from its configuration parameters. This scaling factor is a pre-set decimal between 0 and 1, typically ranging from 0.1 to 0.5, used to control the magnitude of each model update. The cumulative plastic deformation component, verified for reliability, is attenuated according to this scaling factor to generate attenuated nodal displacements. This process can be described by the following formula: ; In this formula, i represents any vertex in the 3D model. α is the cumulative plastic deformation component vector of vertex i. α is the obtained scaling factor. This is the calculated decayed node displacement vector of vertex i. This decayed node displacement vector is applied to the corresponding vertex of the current 3D reference model to update its 3D geometric coordinate data, thereby generating the updated 3D reference model. The vertex coordinate update formula is: ; in, It is the coordinate vector of vertex i in the model before the update. This is the new coordinate vector of vertex i in the updated model. By performing this operation on all vertices of the model, an updated 3D reference model with a shape closer to the current stable state of the physical entity is generated, and this model is used as the starting benchmark for the next monitoring cycle.

[0052] For example, after triggering a model update, a decay update strategy is employed to avoid abrupt changes. A preset scaling factor is used. This means that each update only accepts 20% of the calculated values. Assuming a vertex... Current coordinates Verified cumulative plastic deformation vector That is, a cumulative downward deflection of 5mm. According to the formula... According to the formula have: ; ; Updated vertex coordinates become The system saves the new model. If the actual plastic deformation does indeed stabilize at 5mm, the system will update the model approximately 5-10 times in subsequent update cycles (depending on the triggering frequency). This gradual update approach approximates the true state. Similar to low-pass filtering, this method effectively prevents drastic model jumps caused by single false alarms, ensuring the stability of subsequent differential monitoring.

[0053] Reference Figure 4Based on the same inventive concept, the present invention also provides an intelligent monitoring system for bridge formwork deformation that integrates machine vision, the system comprising: The benchmark construction module is used to obtain a 3D model to describe the geometry of the bridge template and calibration parameters to describe the camera imaging relationship. Based on the 3D model and calibration parameters, a theoretical edge feature map is generated. A distance transformation is performed on the theoretical edge feature map to generate a two-dimensional distance potential energy field. The feature extraction module is used to extract the current video frame from the video stream captured by the on-site camera, and to extract the measured edge feature map from the current video frame; The deformation calculation module is used to construct and solve the inverse projection objective function based on the measured edge feature map and the two-dimensional distance potential energy field, and generate the total displacement field. The decoupling analysis module is used to perform orthogonal decomposition on the total displacement field to decouple the rigid body motion component, the cumulative plastic deformation component and the local elastic deformation component. The output warning module is used to generate deformation visualization results and safety warning signals based on the local elastic deformation component and the cumulative plastic deformation component.

[0054] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0055] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for intelligent monitoring of bridge formwork deformation integrating machine vision, characterized in that, The method includes: A three-dimensional model describing the geometry of the bridge template and calibration parameters describing the camera imaging relationship are obtained. A theoretical edge feature map is generated based on the three-dimensional model and calibration parameters. A distance transformation is performed on the theoretical edge feature map to generate a two-dimensional distance potential energy field. Extract the current video frame from the video stream captured by the on-site camera, and extract the measured edge feature map from the current video frame; Based on the measured edge feature map and the two-dimensional distance potential field, an inverse projection objective function is constructed and solved to generate the total displacement field; Perform orthogonal decomposition on the total displacement field to decouple the rigid body motion component, the cumulative plastic deformation component, and the local elastic deformation component; Based on the local elastic deformation component and the cumulative plastic deformation component, deformation visualization results and safety warning signals are generated.

2. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 1, characterized in that, The generated theoretical edge feature map includes: Obtain the initial deformation field for simulating the minute deformation of bridge templates under initial stress; The initial deformation field is applied to the three-dimensional model to generate a set of candidate three-dimensional models; Each candidate 3D model in the candidate 3D model set is rendered according to the calibration parameters to generate a candidate baseline feature map set; Extract the first frame image of the video stream and process the first frame image to generate the first frame edge feature map; Calculate the image quality confidence weights for each region of the first frame edge feature map, and generate a comprehensive matching score based on the matching degree between the image quality confidence weights and the first frame edge feature map and the candidate baseline feature map set. Based on the overall matching degree, the feature map with the highest matching degree is selected from the candidate benchmark feature map set as the theoretical edge feature map.

3. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 2, characterized in that, The generated total displacement field includes: The initial state of the 3D model is determined based on the theoretical edge feature map, and the vertex displacements are initialized. An edge feature map is generated by projecting the current vertex displacement onto the calibration parameters. The calculated edge feature map is compared with the measured edge feature map to generate a residual map; Analyze the spatial distribution pattern of the residual map, and dynamically adjust the regional weights of the two-dimensional distance potential field according to the spatial distribution pattern to generate an adaptive distance potential field; Based on the physical constraints of the adaptive distance potential energy field and the smoothness of the adjacent vertex displacements, the vertex displacements are iteratively optimized to generate the total displacement field.

4. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 3, characterized in that, The generation of the adaptive distance potential field includes: Identify the distribution pattern of pixels in the residual image and classify it into random noise pattern, directional stripe pattern, or structural continuous pattern; Obtain the image regions corresponding to the random noise pattern and the directional stripe pattern, and generate a weighting factor to reduce the influence of the image regions in the two-dimensional distance potential field. The weighting factor is applied to the two-dimensional distance potential field to generate an adaptive distance potential field.

5. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 3, characterized in that, The method further includes: The convergence speed of the iterative optimization vertex displacement process is obtained, and the confidence score of the optimization result of the current frame is calculated by combining the spatial distribution pattern of the residual map and the image quality confidence weight. The confidence level of the optimization result is correlated with the total displacement field to generate displacement data with confidence level labels.

6. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 1, characterized in that, The orthogonal decomposition of the total displacement field includes: Multiple total displacement fields generated from consecutive video frames are acquired to construct displacement time series for each vertex of the three-dimensional model; The displacement time series was subjected to time-frequency analysis and decomposed into high-frequency displacement components, mid-to-low frequency displacement components, and cumulative plastic deformation components. The time points of vibration events generated by external construction equipment are obtained, and transient interference components are filtered out by correlating the vibration event time points with the occurrence time of high-frequency displacement components. The spatial distribution consistency of low- and medium-frequency displacement components was analyzed, and rigid displacement components and local elastic deformation components were separated.

7. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 1, characterized in that, The generated deformation visualization results and safety warning signals include: A net deformation field is generated by vector synthesis of the local elastic deformation component and the cumulative plastic deformation component. Obtain the safe deformation threshold used to define the deformation hazard level; The net deformation field is compared with the safe deformation threshold to trigger a safety warning signal, and the net deformation field is rendered onto the surface of the three-dimensional model to generate the deformation visualization result.

8. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 7, characterized in that, The method further includes: Assess the overall reliability of the net deformation field; A confidence threshold is obtained to determine whether the result is reliable. When the overall confidence level is higher than the confidence threshold, the cumulative plastic deformation component is back-fused into the three-dimensional model to generate an updated three-dimensional reference model. The updated three-dimensional reference model is then used to perform subsequent monitoring procedures.

9. The intelligent monitoring method for bridge formwork deformation integrating machine vision according to claim 8, characterized in that, The generated updated 3D reference model includes: Obtain the scaling factor used to smooth the model update process; The cumulative plastic deformation component is attenuated according to a proportionality factor to generate attenuated nodal displacements. The decayed node displacements are applied to the corresponding vertices of the 3D model to update the geometric data of the 3D model, generating an updated 3D reference model.

10. A bridge formwork deformation intelligent monitoring system integrating machine vision, applied to the bridge formwork deformation intelligent monitoring method integrating machine vision as described in any one of claims 1-9, characterized in that, The system includes: The benchmark construction module is used to obtain a 3D model to describe the geometry of the bridge template and calibration parameters to describe the camera imaging relationship. Based on the 3D model and calibration parameters, a theoretical edge feature map is generated. A distance transformation is performed on the theoretical edge feature map to generate a two-dimensional distance potential energy field. The feature extraction module is used to extract the current video frame from the video stream captured by the on-site camera, and to extract the measured edge feature map from the current video frame; The deformation calculation module is used to construct and solve the inverse projection objective function based on the measured edge feature map and the two-dimensional distance potential energy field, and generate the total displacement field. The decoupling analysis module is used to perform orthogonal decomposition on the total displacement field to decouple the rigid body motion component, the cumulative plastic deformation component and the local elastic deformation component. The output warning module is used to generate deformation visualization results and safety warning signals based on the local elastic deformation component and the cumulative plastic deformation component.

Images