Concrete column axial deformation recognition method and system based on visual image

Abstract

The application discloses a concrete column axial deformation recognition method and system based on visual images, and relates to the technical field of building structure monitoring.The method comprises the following steps: collecting continuous image sequences of a concrete column from different sides; performing edge detection and morphological processing on the continuous image sequences, extracting the overall contour of the column body, fitting the axial center line of the column body, and generating a global axial length change feature; performing adjacent frame tracking recognition of local evaluation center points to establish a first local deformation feature; performing local deformation feature extraction of a convolutional neural network to establish a second local deformation feature; and performing hierarchical fusion to establish an axial deformation recognition result.The application solves the technical problem that the single feature extraction of the axial deformation of the concrete column in the prior art leads to one-sided deformation information, insufficient monitoring accuracy, and poor reliability, and achieves the technical effects of realizing multi-source feature fusion recognition of the axial deformation of the concrete column and improving the accuracy and reliability of deformation monitoring.

Description

Technical Field

[0001] This invention relates to the field of building structure monitoring technology, specifically to a method and system for identifying axial deformation of concrete columns based on visual images. Background Technology

[0002] In the field of building structural health monitoring, concrete columns, as core load-bearing components, are crucial for structural safety due to their axial deformation monitoring. Current mainstream monitoring technologies have significant limitations: contact sensors (such as strain gauges and displacement meters) require attachment to the column, making installation cumbersome and susceptible to environmental interference; single-point monitoring cannot cover overall deformation. Traditional visual monitoring often uses single-view acquisition and single feature extraction, making it susceptible to occlusion and perspective errors, unable to simultaneously acquire global trends and subtle local displacements, resulting in insufficient monitoring accuracy. Furthermore, the lack of a multi-source deformation information fusion mechanism makes it difficult to integrate global and local, temporal and spatial features, limiting the accuracy of deformation identification and failing to provide reliable support for early warning and accurate assessment of structural risks.

[0003] Existing technologies suffer from the problem of using a single feature to extract axial deformation of concrete columns, resulting in one-sided deformation information, insufficient monitoring accuracy, and poor reliability. Summary of the Invention

[0004] This application provides a method and system for identifying axial deformation of concrete columns based on visual images, which addresses the technical problem that existing technologies use a single feature to extract axial deformation of concrete columns, resulting in one-sided deformation information, insufficient monitoring accuracy, and poor reliability.

[0005] In view of the above problems, this application provides a method and system for identifying axial deformation of concrete columns based on visual images.

[0006] The first aspect of this application provides a method for identifying axial deformation of concrete columns based on visual images, the method comprising:

[0007] A continuous image sequence of a concrete column is acquired from different sides using at least two RGB-D cameras, with the interval angle between the RGB-D cameras being greater than or equal to 45 degrees. Edge detection and morphological processing are performed on the continuous image sequence to extract the overall outline of the column and fit its axial centerline. Global axial length variation features are generated using the time-series identifier of the continuous image sequence and the fitted axial centerline. Preprocessing of the continuous image sequence is performed, and after adaptively selecting a local evaluation center point, adjacent frame tracking and recognition of the local evaluation center point are performed to establish a first local deformation feature. The preprocessed continuous image sequence is then subjected to a convolutional neural network for local deformation feature extraction to establish a second local deformation feature. The global axial length variation feature, the first local deformation feature, and the second local deformation feature are then fused layer by layer to establish an axial deformation recognition result.

[0008] A second aspect of this application provides a system for identifying axial deformation of concrete columns based on visual images, the system comprising:

[0009] The system includes an image sequence acquisition module for acquiring continuous image sequences of a concrete column from different sides using at least two RGB-D cameras, wherein the interval angle between the RGB-D cameras is greater than or equal to 45 degrees; a variation feature generation module for performing edge detection and morphological processing on the continuous image sequence, extracting the overall outline of the column, fitting the axial centerline of the column, and generating global axial length variation features using the time series identifier of the continuous image sequence and the fitted axial centerline of the column; a first local deformation feature establishment module for performing continuous image sequence preprocessing, adaptively selecting a local evaluation center point, and then performing adjacent frame tracking and recognition of the local evaluation center point to establish a first local deformation feature; a second local deformation feature establishment module for extracting local deformation features from the preprocessed continuous image sequence using a convolutional neural network to establish a second local deformation feature; and a recognition result establishment module for performing layered fusion of the global axial length variation feature, the first local deformation feature, and the second local deformation feature to establish an axial deformation recognition result.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] A continuous image sequence of a concrete column is acquired from different sides using at least two RGB-D cameras. Edge detection and morphological processing are performed on the continuous image sequence to extract the overall outline of the column and fit its axial centerline. Global axial length variation features are generated using the time-series identifier of the continuous image sequence and the fitted axial centerline. Preprocessing of the continuous image sequence is performed, and after adaptively selecting a local evaluation center point, adjacent frame tracking and recognition are performed to establish a first local deformation feature. The preprocessed continuous image sequence is then subjected to a convolutional neural network for local deformation feature extraction to establish a second local deformation feature. Layered fusion is then performed to establish the axial deformation recognition result. This method achieves multi-source feature fusion recognition of the axial deformation of concrete columns, improving the accuracy and reliability of deformation monitoring. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1A schematic flowchart of a method for identifying axial deformation of concrete columns based on visual images, provided in an embodiment of this application;

[0014] Figure 2 A schematic diagram of the structure of a visual image-based axial deformation recognition system for concrete columns provided in this application embodiment.

[0015] Figure labeling: Image sequence acquisition module 10, change feature generation module 20, first local deformation feature establishment module 30, second local deformation feature establishment module 40, recognition result establishment module 50. Detailed Implementation

[0016] This application provides a method and system for identifying axial deformation of concrete columns based on visual images, which addresses the technical problem that existing technologies use a single feature to extract axial deformation of concrete columns, resulting in one-sided deformation information, insufficient monitoring accuracy, and poor reliability.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, this application provides a method for identifying axial deformation of concrete columns based on visual images, the method comprising:

[0019] Step S100: Use at least two RGB-D cameras to acquire a continuous image sequence of the concrete column from different sides, wherein the interval angle between the RGB-D cameras is greater than or equal to 45 degrees.

[0020] Specifically, firstly, based on the height, cross-sectional dimensions, and site environment of the concrete column, at least two RGB-D cameras are deployed, ensuring that the cameras capture images from different sides of the concrete column, with an angle of at least 45 degrees between any two cameras to cover key perspectives along the column's circumference and avoid blind spots. Then, all RGB-D cameras are activated to simultaneously acquire image data of the concrete column at different time points, forming a continuous image sequence containing temporal information, providing fundamental data support for subsequent deformation feature analysis.

[0021] Step S200: Perform edge detection and morphological processing on the continuous image sequence to extract the overall contour of the cylinder and fit the axial centerline of the cylinder. Use the time series identifier of the continuous image sequence and the fitted axial centerline of the cylinder to generate global axial length variation features.

[0022] Specifically, edge detection is first performed on each frame of the continuous image sequence. The Canny edge detection algorithm is used to identify the boundary pixels between the concrete column and the background in the image. Then, morphological opening and closing operations are used to denoise the edge detection results, removing noise interference and small redundant contours in the image, and accurately extracting the overall outer contour of the concrete column in each frame. Next, for the outer contour of the column captured by different cameras at the same time point, spatial calibration is performed using the principle of multi-view geometric projection to construct a spatially consistent contour, ensuring that the contour information from different perspectives accurately corresponds to the actual spatial shape of the column. Based on this, straight lines are fitted to the boundary points on both sides of the spatially consistent contour, and the midline of the fitted straight lines on both sides is taken as the axial centerline of the concrete column. Finally, based on the time sequence identifier of the continuous image sequence, the length of the axial centerline is calculated frame by frame. By comparing the differences in the centerline length at different time frames, a global axial length change feature reflecting the overall deformation trend of the column is generated.

[0023] Step S300: Perform continuous image sequence preprocessing, adaptively select local evaluation center points, perform adjacent frame tracking and recognition of local evaluation center points, and establish the first local deformation features.

[0024] Specifically, the continuous image sequence is preprocessed, such as with grayscale correction and image enhancement, to eliminate the impact of factors such as lighting changes and camera shake on image quality. Then, adaptive selection of local evaluation center points is performed: first, concrete column texture recognition is performed on the preprocessed first frame image. If the texture recognition result covers a preset threshold (e.g., 80%) on the column surface, texture evaluation center points are evenly distributed based on areas with obvious texture features, while grid evaluation center points are configured for areas not covered by texture. If the texture coverage does not reach the threshold, grid evaluation center points are first configured on the column image according to a preset grid density, and then, using the grid center points as search centers, surrounding texture key points are searched and replaced with texture key points, ultimately forming a complete set of local evaluation center points. After configuring the center point, optical flow tracking is performed on each local evaluation center point in adjacent frame images. Simultaneously, the tracking results are calibrated to achieve center point alignment by combining local texture gradient field features, depth point cloud constraints acquired by the RGB-D camera, and mutual information similarity matching. Then, a sliding time window is established for the continuous multi-frame trajectory of each center point, and the temporal displacement features of the center point within the window are calculated. Using the center point as the search zero point, a calibration neighborhood range is set. After correcting the neighborhood range by combining neighborhood complexity and temporal displacement features, the local strain tensor, volumetric strain, and shear strain energy density of the neighborhood pixel blocks are extracted to construct the neighborhood strain field. Finally, the consistency of the temporal displacement features is checked through the neighborhood strain field, abnormal displacement data is removed, and a first local deformation feature that can reflect the subtle local deformation of the column is established.

[0025] Step S400: Extract local deformation features from the preprocessed continuous image sequence using a convolutional neural network to establish a second local deformation feature.

[0026] Specifically, a second local deformation feature is extracted using a convolutional neural network. Local image patches containing multiple adjacent frames are extracted from the preprocessed continuous image sequence. The patch size is set according to the detail density of the cylinder surface. These patches are input in a multi-channel format to a U-Net-based residual encoder-decoder network. The encoder of this network extracts multi-scale features of the image through convolution and pooling operations, while the decoder restores the feature map resolution through upsampling and skip connections. Simultaneously, the residual structure effectively alleviates the gradient vanishing problem, improving feature extraction accuracy. After the network runs, it outputs the deformation amplitude (reflecting the magnitude of deformation at the pixel location) and confidence value (reflecting the reliability of the deformation detection result) for each pixel in each frame. Based on these pixel-level deformation data, a second local deformation feature covering the cylinder surface is formed, further supplementing the local deformation information of the cylinder.

[0027] Step S500: The global axial length variation feature, the first local deformation feature, and the second local deformation feature are fused in layers to establish the axial deformation identification result.

[0028] Specifically, the first and second local deformation features are first fused using a weighted average, with the weights dynamically adjusted based on the confidence values ​​of the two types of features. This integrates the deformation data of the two types of local features to obtain a unified local deformation fusion feature. Then, the global axial length variation feature is fused with the local deformation fusion feature a second time. Combined with the structural characteristics of the concrete column (such as the axial stiffness distribution of the column), a global-local correlation mapping is established to ensure that the overall deformation trend and local minute deformations can corroborate each other. After fusion, the axial deformation identification result of the concrete column is generated. This result includes both the overall axial length variation data of the column and the details of local deformation in different areas. Based on the identification result, the overall deformation amplitude and the maximum deformation rate per unit time of the column can be further identified, establishing a first deformation warning and a second deformation warning respectively. When the deformation data exceeds a preset safety threshold, an abnormal alarm is triggered. Furthermore, the deformation monitoring cycle is reconstructed based on the deformation pattern, providing a basis for the long-term monitoring and management of concrete columns.

[0029] In one possible implementation, step S300 further includes:

[0030] Step S310: Obtain the first image in the continuous image sequence, perform concrete column texture recognition on the first image, and establish texture recognition results.

[0031] Step S320: If the coverage of the texture recognition result meets the preset coverage threshold, then the texture recognition result is used as the main recognition result. After adaptively distributing the texture evaluation center points, the concrete columns that do not cover the texture evaluation center points are configured with grid evaluation center points.

[0032] Step S330: Establish local evaluation center points from the mesh evaluation center point and texture evaluation center point to perform adjacent frame tracking and recognition, and establish the first local deformation feature.

[0033] Specifically, the first frame of a preprocessed continuous image sequence is selected as the first image. The concrete column region is extracted from this first image using an image cropping tool to eliminate background interference. Then, a texture feature extraction technique based on the gray-level co-occurrence matrix (GLCM) is employed to segment the cropped column region into blocks. The GLCM of each image block is calculated in four directions (0°, 45°, 90°, and 135°), extracting four texture feature parameters: contrast, correlation, energy, and entropy. This quantifies the distribution pattern and detail differences of the column surface texture. Simultaneously, a local binary pattern (LBP) algorithm is used to perform neighborhood gray-level comparisons on the pixels in the column region, generating a LBP histogram to further capture the local structural features of the texture. The texture features extracted by the two techniques are then fused. A threshold segmentation algorithm is used to process the fused feature data, marking regions with obvious texture features. A region growing algorithm is then used to optimize the boundaries of the marked regions. Finally, the location information, size, and feature intensity data of each texture region are integrated to form a complete texture recognition result.

[0034] The coverage of the obtained texture recognition result is calculated, which is the proportion of the area of ​​the texture recognition marked region in the entire concrete column image area. This proportion is then compared with a preset coverage threshold. If the coverage reaches the preset coverage threshold, it indicates that most areas of the column surface can be accurately located based on texture features, and the texture recognition result is used as the primary recognition basis. Subsequently, following the principle that the more significant the texture features, the higher the distribution density of evaluation center points, texture evaluation center points are adaptively distributed within the texture recognition marked area to ensure that areas with rich texture details can capture local deformation information through more center points. At the same time, for local areas of the concrete column not covered by texture recognition, grid evaluation center points are uniformly configured according to a preset grid density to ensure that the evaluation center points can completely cover the entire concrete column image area, providing a comprehensive localization basis for subsequent tracking and recognition of adjacent frames.

[0035] The coordinates of the previously obtained grid evaluation center points and texture evaluation center points are uniformly calibrated to transform the positional information of the two types of center points to the same image coordinate system. Simultaneously, abnormal center points that may have overlapping coordinates or exceed the imaging range of the column are eliminated, ensuring that both types of center points can completely and without redundancy cover the entire imaging area of ​​the concrete column, forming a local evaluation center point set containing all valid evaluation points. Next, using this local evaluation center point set as the tracking object, tracking and recognition are performed in adjacent frames of a continuous image sequence: the positional changes of each local evaluation center point between adjacent frames are captured using an optical flow tracking algorithm. Combined with local texture gradient field feature verification, RGB-D depth point cloud data constraints, and mutual information similarity matching, possible offsets and losses during the tracking process are calibrated and supplemented to achieve accurate alignment of each center point between adjacent frames. Subsequently, a sliding time window was set for the continuous multi-frame displacement trajectory of each local evaluation center point. The temporal displacement characteristics of the center point within the window were calculated. Simultaneously, using each center point as the search zero point, the features of its neighboring pixel blocks were extracted and a neighborhood strain field was constructed, including the local strain tensor, volumetric strain, and shear strain energy density. The consistency of the temporal displacement characteristics was verified using the neighborhood strain field to eliminate false displacement data caused by noise and illumination changes. Finally, the temporal displacement characteristics of all verified local evaluation center points and the neighborhood strain field information were integrated to establish a first local deformation feature that can accurately reflect the local subtle deformation state of the concrete column surface.

[0036] In one possible implementation, step S310 further includes:

[0037] Step S311: If the coverage of the texture recognition result does not meet the preset coverage threshold, then a grid center point distribution instruction is generated.

[0038] Step S312: Adaptively execute the grid evaluation center point configuration of the concrete column according to the grid center point distribution instruction.

[0039] Step S313: Using the configured mesh evaluation center point as the search center, perform a texture key point association search based on the texture recognition results.

[0040] Step S314: After replacing the corresponding grid evaluation center point using correlation search, the construction of local evaluation center points is completed.

[0041] Specifically, the coverage of the concrete column texture recognition result in the first image of a continuous image sequence is calculated, specifically by statistically analyzing the proportion of the texture recognition area within the overall imaging area of ​​the concrete column. This proportion is then compared to a preset coverage threshold. If the calculated coverage does not reach the preset threshold, it indicates that the current texture recognition result cannot fully cover the concrete column imaging area, making it difficult to support the accurate extraction of subsequent local deformation features. At this point, an automatic instruction generation mechanism is triggered to generate a grid center point distribution instruction to guide the subsequent configuration of evaluation center points, providing an operational basis for supplementing evaluation center points and filling monitoring gaps not covered by texture recognition through grid division.

[0042] The core parameters contained in the grid center point distribution command are analyzed to clarify key information related to concrete column imaging, including the actual size of the concrete column in the image, image resolution, and preset deformation monitoring accuracy requirements. Subsequently, based on this information, the specific specifications for grid division are determined: an appropriate grid density is planned according to the height and width of the concrete column imaging area to ensure that the divided grid meets the monitoring accuracy requirements without causing data redundancy due to excessive grid density. During the division process, for irregular imaging areas at the edges of the concrete column, the size and shape of the local grid are adjusted to avoid monitoring blind spots or wasted evaluation center points in edge areas due to a fixed grid pattern. Finally, according to the completed grid structure, a corresponding grid evaluation center point is generated at the geometric center of each grid cell, ensuring that all grid evaluation center points uniformly and completely cover the entire imaging area of ​​the concrete column, providing comprehensive monitoring markers for subsequent local deformation feature extraction.

[0043] The specific locations of all configured grid evaluation center points in the first image of the concrete column are determined, and the pixel coordinate range corresponding to each grid evaluation center point is defined. Next, feature information of all texture key points is extracted from the previously obtained texture recognition results. This feature information includes the gray-level distribution characteristics, local texture structure characteristics, and pixel value variation patterns of the texture key points. Subsequently, for each grid evaluation center point, a fixed search area is delineated with that center point as the core, ensuring that the search area covers the image range around the center point where texture key points may exist. Within the delineated search area, feature extraction is performed on each image pixel within the area to obtain the texture feature data of each pixel. This data is then compared one by one with the texture key point feature information extracted from the texture recognition results. During the comparison, a feature similarity calculation method is used to determine the similarity between the pixel features within the search area and the texture key point features. If the similarity between a pixel feature and a texture key point feature reaches a preset standard, the grid evaluation center point and the texture key point are successfully associated; if no pixel feature within the search area meets the similarity standard, no association is established. Following the above method, perform the associated search operation of texture key points for all mesh evaluation center points in sequence.

[0044] The search results associated with all grid evaluation center points were systematically reviewed to identify those successfully associated with texture keypoints. These were grid evaluation center points whose features, through feature comparison, met the similarity criteria with the texture keypoints in the texture recognition results. For these grid evaluation center points, their original positions were directly replaced with their associated texture keypoints. This leverages the advantages of texture keypoints—their closer fit to the concrete column surface texture and higher positioning accuracy—to improve the accuracy of subsequent deformation tracking. For grid evaluation center points that were not successfully associated with texture keypoints, their original grid center positions were retained because there was no effective texture information available for reuse in their areas, ensuring no monitoring gaps in the concrete column imaging area. Finally, all replaced texture keypoints and retained grid evaluation center points were integrated to form a point set covering the entire imaging area of ​​the concrete column, possessing both texture feature fit and positional integrity, thus completing the construction of local evaluation center points.

[0045] In one possible implementation, step S300 further includes:

[0046] Step S340: Perform optical flow tracking of the local evaluation center point in adjacent frames, and perform tracking calibration through local texture gradient field features, RGB-D depth point cloud constraints and mutual information similarity matching to perform local evaluation center point alignment.

[0047] Step S350: Establish a sliding time window for the continuous multi-frame trajectory of each local evaluation center point, and calculate the temporal displacement characteristics of the local evaluation center point based on the sliding time window.

[0048] Step S360: Using the local evaluation center point as the search zero point, perform feature extraction of the neighboring pixel block and establish a neighborhood strain field, which includes the local strain tensor, volumetric strain, and shear strain energy density.

[0049] Step S370: After verifying the consistency of the time-series displacement characteristics using the neighborhood strain field, establish the first local deformation characteristics.

[0050] Specifically, preprocessing is performed on two adjacent frames in a continuous image sequence to unify the image resolution and grayscale value range, laying the foundation for optical flow tracking. Then, the Lucas-Kanade optical flow algorithm is employed, using all local evaluation center points in the reference frame as initial feature points, to search for corresponding points with similar grayscale distributions in the target frame. By calculating the grayscale gradient of the 3×3 or 5×5 pixel neighborhood around each local evaluation center point, an optical flow equation is established, and the initial displacement vector of the center point in the target frame is obtained, thus achieving initial optical flow tracking. To correct potential deviations in initial tracking, a triple calibration method is introduced: First, the local texture gradient field features of a 10×10 pixel region surrounding each local evaluation center point in the reference frame and target frame are extracted. By calculating the cosine similarity of the gradient direction histogram, tracking results with gradient direction consistency higher than a preset standard are selected, eliminating incorrect matches caused by texture changes. Second, depth data collected by an RGB-D camera is used to generate a 3D point cloud model of the concrete column. The 2D pixel coordinates of the local evaluation center points in the reference frame and target frame are converted into 3D spatial coordinates. By calculating the spatial distance between corresponding points in the two frames, the 3D coordinate deviation is constrained within a set range, eliminating tracking errors caused by perspective distortion of the 2D image. Third, the mutual information value of a 15×15 pixel region centered on the local evaluation center point in the reference frame and target frame is calculated. A higher mutual information value indicates a stronger correlation between the two regions. This is used as a basis for judging the effectiveness of the tracking results, retaining tracking points with mutual information values ​​higher than a threshold. Finally, by combining the triple calibration results, the position of the target frame center point obtained from the initial optical flow tracking is corrected, the precise coordinates of each local evaluation center point in the target frame are determined, the alignment of local evaluation center points between adjacent frames is completed, and the consistency of the position reference for subsequent deformation calculation is ensured.

[0051] For each local evaluation center point, complete displacement trajectory data is extracted from its continuous multi-frame tracking records. This data includes the axial and radial coordinate changes of the center point in different frames. Subsequently, a sliding time window is configured for the displacement trajectory of each center point. The window size is set according to the frame rate of the image sequence and the temporal resolution requirements of deformation monitoring, such as covering 5-10 consecutive frames. The window can slide frame by frame along the time axis to ensure dynamic capture of displacement changes within different time periods. Within each sliding time window, the displacement data of the frames covered by the window are quantitatively calculated: the maximum, minimum, and average axial displacement of the center point within the window are statistically analyzed to determine the range of displacement changes; the difference in displacement between adjacent frames is calculated to obtain the displacement change rate; and a linear fitting algorithm is used to determine the overall trend of displacement change within the window, such as uniform deformation, accelerated deformation, or stable deformation with no significant deformation. These calculated displacement amplitudes, change rates, and trend characteristics are integrated to form the temporal displacement characteristics of each local evaluation center point within the corresponding time window. This achieves accurate characterization of the dynamic process of center point displacement, providing a temporal displacement basis for subsequent deformation consistency verification and the establishment of the first local deformation characteristics.

[0052] Each local assessment center point is set as the search zero point. Based on the imaging resolution and deformation monitoring accuracy requirements of the concrete column, a fixed neighborhood range centered on the search zero point is defined, such as a 15×15 pixel area. This neighborhood is then uniformly divided into several independent pixel blocks to ensure comprehensive coverage of the area around the search zero point where deformation may occur. Next, feature extraction is performed on each pixel block: the grayscale values, two-dimensional coordinates, and RGB-D depth data of all pixels within each pixel block are obtained. The center coordinates and mean grayscale value of the pixel block are calculated, and the relative positional relationship and grayscale difference between adjacent pixel blocks are recorded to provide basic data for subsequent strain calculations. Subsequently, a neighborhood strain field is constructed based on the extracted pixel block features: by calculating the axial and radial displacement changes of the pixel blocks within the neighborhood, and combining this with an elastic strain calculation model, a local strain tensor is derived. This tensor includes: normal strain, reflecting the degree of axial and radial tension or compression; and shear strain, reflecting the relative shear deformation between adjacent pixel blocks, which can quantify the deformation state in different directions within the neighborhood. Based on the volume change rate of each pixel block within the neighborhood, derived from depth data and two-dimensional coordinate changes, the volumetric strain is calculated to characterize the overall volume expansion or contraction of the neighborhood region. Then, based on the shear strain component in the local strain tensor, combined with the elastic parameters of concrete materials, such as shear modulus, the shear strain energy density is solved using an energy calculation formula, reflecting the energy level accumulated by shear deformation within the neighborhood. Finally, by integrating the local strain tensor, volumetric strain, and shear strain energy density, a neighborhood strain field corresponding to each local evaluation center point is formed, comprehensively characterizing the microscopic deformation features of the area surrounding the center point, providing a spatial strain basis for subsequent consistency verification of temporal displacement features.

[0053] For each local evaluation center point, a correlation analysis is performed between its corresponding established neighborhood strain field and the obtained temporal displacement features: The deformation direction and magnitude reflected by the neighborhood strain tensor are compared with the displacement trend and amplitude in the temporal displacement features to verify whether the overall deformation state reflected by volumetric strain matches the cumulative displacement effect, and the distribution law of shear strain energy density and the drastic active gradient of displacement change correspond. If contradictions arise between the two types of parameters during the verification process, such as the temporal strain tensor showing compression towards compression while the temporal displacement features show axial stretching, then the corresponding temporal displacement features are determined to be abnormal. Correction is required based on the feature distribution of the neighborhood pixel blocks, eliminating erroneous displacement data affected by noise or local interference. After the temporal displacement features of all local evaluation center points pass the consistency verification, the verified temporal displacement features of each center point are integrated with the corresponding neighborhood strain field information to form a set that accurately reflects the dynamic deformation process in the time dimension and the deformation distribution characteristics in the spatial dimension of different areas on the concrete column surface. This set is used to establish the first local deformation feature.

[0054] In one possible implementation, step S360 further includes:

[0055] Step S361: Set the calibration neighborhood range. After configuring the initial neighborhood region based on the search zero point and the calibration neighborhood range, perform neighborhood complexity calculation within the initial neighborhood region and establish the first compensation factor.

[0056] Step S362: Obtain the temporal displacement features corresponding to the search zero point, and establish a second compensation factor based on the temporal displacement features.

[0057] Step S363: After performing calibration neighborhood range correction using the first compensation factor and the second compensation factor, perform feature extraction of the neighborhood pixel block.

[0058] Specifically, based on the typical scale of concrete column surface texture and the accuracy requirements of deformation monitoring, a fixed-size calibration neighborhood is preset, such as a rectangular area in pixels, to ensure that this range can initially cover the basic area around the search zero point where significant deformation information may exist. Next, using each local evaluation center point as the search zero point, the calibration neighborhood is spatially located around the search zero point to form an initial neighborhood region, clarifying the set of pixels covered by the region boundary. Subsequently, multi-dimensional neighborhood complexity calculations are performed on the initial neighborhood region: through gray-level histogram analysis, the standard deviation and range of pixel gray-level values ​​within the region are calculated to quantify the dispersion of gray-level distribution; the Sobel operator is used to extract texture edges within the region, and the proportion of edge pixels to the total number of pixels in the region is statistically analyzed to assess the density of the texture structure; the sliding window method is used to divide the region into several sub-blocks, and the difference coefficient of gray-level mean between sub-blocks is calculated to measure the non-uniformity of local texture; Fourier transform is used to analyze the frequency components of the texture within the region to determine the proportion of high-frequency components to reflect the richness of detail. The aforementioned grayscale dispersion, edge density, sub-block difference coefficient, and high-frequency component proportion are weighted and fused to obtain a comprehensive quantitative neighborhood complexity value. A higher value indicates a more complex texture and richer features within the region, while a lower value indicates a simpler texture and smoother features. Finally, a first compensation factor is established based on this neighborhood complexity value. The complexity value and the compensation factor are positively correlated, thereby quantitatively representing the extent to which the subsequent neighborhood range needs to be adjusted according to the texture complexity.

[0059] The system retrieves the temporal displacement feature dataset corresponding to each search zero point from the stored time-series data. This dataset contains the axial coordinate changes, radial coordinate changes, and inter-frame displacement vectors of that point across N consecutive frames. Multi-dimensional calculations are performed on this data: the cumulative displacement is obtained by summing the absolute values ​​of the axial and radial displacements; the displacement fluctuation rate is obtained by calculating the mean of the squared differences in displacement between adjacent frames; and the displacement-frame variation curve is linearly fitted using the least squares method, with the absolute value of the slope of the fitted line taken as the displacement trend strength. These three indicators are then normalized (mapped to the 0-1 range) and weighted according to preset weights, such as 0.4 for cumulative displacement, 0.3 for displacement fluctuation rate, and 0.3 for trend strength, to obtain a comprehensive dynamic index. Finally, this index is substituted into a preset mapping function, such as an exponential function, to generate a second compensation factor. Search zero points with higher comprehensive dynamic indices correspond to larger compensation factor values, thus quantifying the adjustment requirements of the neighborhood range.

[0060] The first and second compensation factors are standardized to map their values ​​to the same order of magnitude. Then, they are weighted according to preset weights (e.g., the first compensation factor accounts for 0.6 and the second compensation factor accounts for 0.4) to obtain a comprehensive correction coefficient. Based on this comprehensive correction coefficient, the initially set calibration neighborhood is dynamically adjusted: when the comprehensive correction coefficient is greater than 1, the neighborhood is expanded proportionally to include more pixel information in areas with high texture complexity or high dynamic deformation; when the coefficient is less than 1, the neighborhood is reduced proportionally to decrease redundant pixels in low-complexity or low-deformation areas. The corrected neighborhood range adapts to the complexity of local textures, matches the monitoring requirements of dynamic deformation, and always maintains a symmetrical distribution centered on the search zero point. Subsequently, pixel block feature extraction is performed within the adjusted neighborhood region: the region is uniformly divided into several non-overlapping pixel sub-blocks, and the gray mean, gray variance, center coordinates, and depth values ​​based on RGB-D data of each sub-block are extracted one by one. At the same time, the gradient direction and intensity of the sub-block edges are calculated, and finally a structured data set containing gray features, spatial coordinate features, and depth features is formed, providing accurate basic feature parameters for the subsequent construction of the neighborhood strain field.

[0061] In one possible implementation, step S400 further includes:

[0062] Step S410: Extract local image patches containing multiple adjacent frames from the continuous image sequence, and input the local image patches into a convolutional neural network through multiple channels. The convolutional neural network is a residual encoder-decoder network based on U-Net, and outputs the deformation amplitude and confidence value of each pixel.

[0063] Step S420: Establish a second local deformation feature based on the deformation amplitude and confidence value of each pixel.

[0064] Specifically, from a continuous image sequence of concrete columns, image segments containing multiple adjacent frames (e.g., 3-5 adjacent frames) are extracted at fixed time intervals. Then, using each local evaluation center point as the core, fixed-size local image patches are cropped to ensure that the patches cover the area around the center point where deformation may occur, and that each patch retains the temporal correlation information of multiple frames. Subsequently, these local image patches are input into a convolutional neural network in a multi-channel format. This network is based on the U-Net architecture and incorporates residual connections: in the encoding stage, low-dimensional to high-dimensional deformation features of the image patches are progressively extracted through convolutional and pooling layers. Residual connections effectively alleviate the gradient vanishing problem in deep network training, ensuring the integrity of feature transfer; in the decoding stage, upsampling and skip connections fuse the features extracted in the encoding stage with shallow detail features, achieving pixel-level precise localization of the image patches. Ultimately, the network outputs two key results for each local image patch: first, the deformation amplitude of each pixel in the axial and radial directions, quantifying the specific degree of deformation at the pixel level; and second, the confidence value corresponding to the deformation amplitude of each pixel, reflecting the reliability of the deformation detection result of that pixel. The higher the confidence value, the stronger the accuracy of the deformation amplitude.

[0065] The pixel deformation amplitude and confidence value of each local image patch output by the convolutional neural network are filtered, and pixel deformation data with confidence values ​​below a preset threshold are removed to avoid low-reliability information interfering with feature accuracy. Next, the filtered pixel deformation amplitude values ​​are mapped to the imaging region coordinates of the concrete column according to the position of the local image patch in the continuous image sequence, forming a pixel-level deformation amplitude distribution map. Simultaneously, the confidence value distribution of high-confidence pixels is retained as a supplementary explanation of the deformation amplitude reliability. Finally, the pixel-level deformation amplitude distribution map and the confidence value distribution are integrated to construct a second local deformation feature that combines pixel-level accuracy and reliability verification. This feature can reflect the deformation differences of different areas on the surface of the concrete column at a microscopic level, providing another dimension of local deformation basis for subsequent overall axial deformation identification.

[0066] In one possible implementation, step S200 further includes:

[0067] Step S210: On each frame of the image, after Canny edge detection, denoising is performed using morphological opening and closing operations to extract the outer contour of the cylinder.

[0068] Step S220: Perform multi-view geometric projection on the outer contour of the cylinder from different perspectives at the same time point to establish a spatially consistent contour.

[0069] Step S230: Perform straight line fitting on both sides of the boundary points based on the spatial consistency profile, and use the midline of the straight line fitting results on both sides as the axial centerline of the cylinder.

[0070] Step S240: Extract global length change features from frame-by-frame images based on the axial centerline of the cylinder to generate global axial length change features.

[0071] Specifically, for each frame in a continuous image sequence, the Canny edge detection algorithm is used for preliminary edge extraction. By setting dual thresholds—a high threshold for identifying strong edges and a low threshold for connecting weak edges—the image grayscale gradient is calculated and judged to filter out pixels with drastic grayscale changes, thus initially forming the edge contour between the concrete column and the background. Since the preliminary detection results may contain fragmented edges or isolated points caused by noise interference, further morphological processing is performed: first, a morphological opening operation is performed, using a preset structuring element to erode the edge image to eliminate fine noise and burrs; then, a dilation operation is performed to restore the main shape of the edge. Next, a morphological closing operation is performed, using a dilation-erosion process to fill in tiny holes and breaks in the edge contour, making the contour lines more continuous and complete. After the above denoising processing, a clear and closed outer contour of the column is separated from the image. This contour accurately surrounds the imaging area of ​​the concrete column in a single frame image, providing a reliable single-frame contour basis for subsequent multi-view contour fusion and axial centerline extraction.

[0072] For images acquired at the same time point from different perspectives (such as front, side, or oblique), the extracted outer contours of the cylinder from each perspective are imported into a multi-view geometric projection model. Using the internal and external parameters obtained from camera calibration, the projection relationship of the contours from different perspectives in three-dimensional space is calculated to eliminate contour deformation caused by perspective differences. Then, through spatial consistency verification, contour areas with high overlap and matching geometric features from each perspective are selected and fused to form a spatially consistent contour that reflects the true spatial form of the cylinder, avoiding occlusion or perspective errors that may exist in single-view contours.

[0073] Based on the established spatially consistent profile, the boundary point sets on both sides of the profile are extracted, including the left boundary point and the right boundary point. The least squares method is used to fit the two boundary point sets to obtain two parallel cylindrical side boundary lines. The vertical distance between the two side boundary lines is calculated, and the trajectory line formed by the midpoint of the distance is taken as the axial center line of the cylinder. This center line can accurately reflect the axial extension direction of the cylinder and provide a benchmark for subsequent length measurement.

[0074] For each frame in a continuous image sequence, based on the determined axial centerline of the cylinder, a contour vertex detection algorithm is used to locate the endpoints of the cylinder boundary on the centerline. The image pixels are traversed along the axial centerline, and the pixel coordinates corresponding to the top and bottom of the cylinder on the centerline are accurately identified by combining the grayscale differences and geometric shapes of the cylinder's outer contour. This ensures that endpoint positioning is not affected by surface texture or local shadows. Next, based on the coordinates of the top and bottom endpoints in each frame, the Euclidean distance between the two points on the axial centerline is calculated. This distance represents the axial length of the cylinder in the corresponding frame. Simultaneously, the timestamp corresponding to the length data of each frame is recorded, establishing a basic time-length mapping relationship. Subsequently, a time-series analysis is performed on the axial length data of the cylinder in consecutive frames: the length difference between adjacent frames is calculated to obtain the instantaneous length change; the average length change per unit time is statistically analyzed to determine the rate of length change; the total length change within the monitoring period is obtained by accumulating the changes over each time period; and the overall trend of length change, such as uniform shortening, accelerated elongation, or phased stabilization, is determined through linear fitting or trend analysis. Finally, the parameters such as frame-by-frame length data, instantaneous change, rate of change, cumulative change, and trend are integrated to form a global axial length change characteristic that can comprehensively reflect the macroscopic axial deformation state of the concrete column.

[0075] In one possible implementation, step S500 further includes:

[0076] Step S510: Perform overall deformation amplitude identification on the axial deformation identification results and establish a first deformation warning.

[0077] Step S520: Identify the maximum deformation rate per unit time based on the axial deformation identification results, and establish a second deformation early warning.

[0078] Step S530: Execute deformation anomaly reporting based on the first deformation warning and the second deformation warning.

[0079] Specifically, key parameters such as the cumulative change in global axial length of the concrete column during the monitoring period and the maximum displacement amplitude of each local evaluation center point are extracted from the axial deformation identification results. The overall deformation amplitude of the column is determined comprehensively. If the cumulative change in global length exceeds the preset safety threshold, which is set according to the design bearing capacity, material properties and usage scenarios of the concrete column, or if the maximum displacement amplitude of a local area exceeds the allowable range, it is determined that the overall deformation of the column has reached the warning standard, and the first deformation warning is established.

[0080] Dynamic parameters such as the rate of change of column axial length per unit time (e.g., per minute) and the displacement rate of local evaluation center point are extracted from the axial deformation identification results. The maximum deformation rate within the monitoring period is calculated. If the maximum deformation rate per unit time exceeds the preset rate threshold, which is based on the creep characteristics of concrete materials and structural stability requirements to avoid structural risks caused by rapid deformation in a short period of time, the column deformation rate is determined to be abnormal, and a second deformation warning is established.

[0081] The triggering status of the first and second deformation warnings is integrated and judged. If only the first deformation warning is triggered, it indicates that the cumulative deformation of the column exceeds the limit but the deformation rate is stable, and a "cumulative deformation anomaly" is reported, suggesting that the impact of long-term deformation on structural performance needs to be assessed. If only the second deformation warning is triggered, it indicates that the deformation rate of the column is too fast but the cumulative amount does not exceed the limit, and a "deformation rate anomaly" is reported, suggesting that the possibility of sudden loads or local damage needs to be investigated. If both warnings are triggered simultaneously, it indicates that the column has both cumulative deformation exceeding the limit and rapid deformation, and a "severe deformation anomaly" is reported, with the warning information being pushed out first. All anomaly reports include the warning type, trigger threshold, actual monitored value, corresponding time and location information, and a visual report is generated to provide accurate basis for formulating structural maintenance or emergency treatment plans.

[0082] In one possible implementation, step S500 further includes:

[0083] Step S540: Reconstruct the deformation monitoring cycle based on the axial deformation identification results.

[0084] Step S550: Use the reconstructed deformation monitoring cycle to monitor and manage concrete columns.

[0085] Specifically, key analytical indicators are extracted from the axial deformation identification results, including the cumulative change in global axial length, the average deformation rate per unit time, the fluctuation range of local deformation characteristics, and the historical triggering of the first and second deformation warnings. These indicators are used to assess the current deformation activity of the column: if the identification results show that the overall deformation of the column is gradual, the cumulative change is far below the safety threshold, and no warnings are triggered for a long period of time, it indicates that the column is in a stable deformation stage, and the original monitoring cycle can be appropriately extended to reduce unnecessary consumption of monitoring resources; if the identification results show that the deformation rate of the column fluctuates significantly, or there is a continuous deformation trend in a local area, or even a deformation warning has been triggered, it indicates that the column is in a dynamic deformation stage, and the original monitoring cycle needs to be shortened and the monitoring frequency increased to accurately capture dynamic changes in deformation; if the deformation characteristics of a local area differ significantly from the overall trend, an independent special monitoring cycle needs to be added for that area in addition to the overall monitoring cycle to achieve enhanced monitoring of key areas. At the same time, by combining the duration of the reconstructed monitoring cycle, the acquisition interval of image data and the time node of deformation analysis are adjusted synchronously, so that the reconstructed monitoring cycle can not only fully cover the time period when the column may undergo significant deformation, but also avoid data redundancy caused by overly dense monitoring, and finally form a dynamic monitoring cycle system that is adapted to the real-time deformation state of the column.

[0086] Work is carried out based on the reconstructed deformation monitoring cycle. First, images of the column are collected at the frequency determined by the new cycle, and axial deformation is identified periodically to keep track of the column's deformation. If abnormal deformation is found or an early warning is triggered within the cycle, the monitoring interval is shortened and the monitoring intensity is increased. At the same time, the dynamic deformation of the column is recorded based on the monitoring data of different cycles, providing a reference for subsequent column maintenance and risk investigation, so as to ensure that the monitoring and management are both efficient and in line with the actual deformation state of the column.

[0087] Example 2 is based on the same inventive concept as the visual image-based axial deformation recognition method for concrete columns in the previous examples, such as... Figure 2 As shown, this application provides a system for recognizing axial deformation of concrete columns based on visual images. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0088] The image sequence acquisition module 10 is used to acquire a continuous image sequence of a concrete column from different sides using at least two RGB-D cameras, wherein the interval angle between the RGB-D cameras is greater than or equal to 45 degrees.

[0089] The variation feature generation module 20 is used to perform edge detection and morphological processing on the continuous image sequence, extract the overall contour of the column, fit the axial center line of the column, and generate global axial length variation features using the time series identifier of the continuous image sequence and the fitted axial center line of the column.

[0090] The first local deformation feature establishment module 30 is used to perform continuous image sequence preprocessing, adaptively select the local evaluation center point, perform adjacent frame tracking and recognition of the local evaluation center point, and establish the first local deformation feature.

[0091] The second local deformation feature establishment module 40 is used to extract local deformation features from the preprocessed continuous image sequence using a convolutional neural network to establish the second local deformation features.

[0092] The recognition result establishment module 50 is used to perform layered fusion of the global axial length change feature, the first local deformation feature, and the second local deformation feature to establish the axial deformation recognition result.

[0093] Furthermore, the system is also used to implement the following functions:

[0094] The first image in a continuous image sequence is acquired, and concrete column texture recognition is performed on the first image to establish a texture recognition result. If the coverage of the texture recognition result meets a preset coverage threshold, the texture recognition result is used as the main recognition result. After adaptively distributing the texture evaluation center points, the concrete columns that do not cover the texture evaluation center points are configured with grid evaluation center points. Local evaluation center points are established from the grid evaluation center points and the texture evaluation center points to perform adjacent frame tracking recognition and establish the first local deformation feature.

[0095] Furthermore, the system is also used to implement the following functions:

[0096] If the coverage of the texture recognition result does not meet the preset coverage threshold, a grid center point distribution instruction is generated; the grid evaluation center point configuration of the concrete column is adaptively executed according to the grid center point distribution instruction; the configured grid evaluation center point is used as the search center to perform texture key point association search of the texture recognition result; after replacing the corresponding grid evaluation center point with the association search, the local evaluation center point construction is completed.

[0097] Furthermore, the system is also used to implement the following functions:

[0098] Optical flow tracking of local evaluation center points is performed in adjacent frames, and tracking calibration is performed through local texture gradient field features, RGB-D depth point cloud constraints, and mutual information similarity matching to align local evaluation center points. A sliding time window is established for the continuous multi-frame trajectory of each local evaluation center point, and the temporal displacement features of the local evaluation center point are calculated based on the sliding time window. Using the local evaluation center point as the search zero point, feature extraction of neighboring pixel blocks is performed to establish a neighborhood strain field, which includes local strain tensor, volumetric strain, and shear strain energy density. After verifying the consistency of the temporal displacement features using the neighborhood strain field, the first local deformation feature is established.

[0099] Furthermore, the system is also used to implement the following functions:

[0100] After setting the calibration neighborhood range, configuring the initial neighborhood region based on the search zero point and the calibration neighborhood range, performing neighborhood complexity calculation within the initial neighborhood region, and establishing a first compensation factor; obtaining the temporal displacement features corresponding to the search zero point, and establishing a second compensation factor based on the temporal displacement features; after correcting the calibration neighborhood range using the first compensation factor and the second compensation factor, performing feature extraction of the neighborhood pixel blocks.

[0101] Furthermore, the system is also used to implement the following functions:

[0102] Local image patches containing multiple adjacent frames are extracted from the continuous image sequence. The local image patches are then input into a convolutional neural network (CNN) via multiple channels. The CNN is a U-Net-based residual encoder-decoder network that outputs the deformation amplitude and confidence value of each pixel. A second local deformation feature is then established based on the deformation amplitude and confidence value of each pixel.

[0103] Furthermore, the system is also used to implement the following functions:

[0104] In each frame of the image, after Canny edge detection, denoising is performed using morphological opening and closing operations to extract the outer contour of the cylinder. The outer contours of the cylinder from different viewpoints at the same time point are subjected to multi-view geometric projection to establish a spatially consistent contour. Based on the spatially consistent contour, straight lines are fitted to the boundary points on both sides, and the midline of the straight line fitting results on both sides is used as the axial centerline of the cylinder. Based on the axial centerline of the cylinder, global length change features are extracted for each frame of the image to generate global axial length change features.

[0105] Furthermore, the system is also used to implement the following functions:

[0106] The overall deformation amplitude of the axial deformation identification result is identified to establish a first deformation warning; the maximum deformation rate per unit time of the axial deformation identification result is identified to establish a second deformation warning; and deformation anomaly reporting is executed based on the first deformation warning and the second deformation warning.

[0107] Furthermore, the system is also used to implement the following functions:

[0108] The deformation monitoring cycle is reconstructed based on the axial deformation identification results; the reconstructed deformation monitoring cycle is used for concrete column monitoring and management.

[0109] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0110] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0111] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for identifying axial deformation of concrete columns based on visual images, characterized in that, The method includes: A continuous image sequence of a concrete column is acquired from different sides using at least two RGB-D cameras, wherein the RGB-D cameras are spaced at an angle greater than or equal to 45 degrees. Edge detection and morphological processing are performed on the continuous image sequence to extract the overall contour of the cylinder and fit the axial centerline of the cylinder. Global axial length variation features are generated using the time series identifier of the continuous image sequence and the fitted axial centerline of the cylinder. After performing continuous image sequence preprocessing and adaptively selecting the local evaluation center point, the adjacent frame tracking and recognition of the local evaluation center point is performed to establish the first local deformation feature. The preprocessed continuous image sequence is subjected to local deformation feature extraction by convolutional neural network to establish a second local deformation feature. The global axial length variation feature, the first local deformation feature, and the second local deformation feature are fused in layers to establish an axial deformation identification result. This includes first fusing the first local deformation feature and the second local deformation feature by weighted averaging, with the weights dynamically adjusted based on the confidence values ​​of the two types of features, and integrating the deformation data of the two types of local features to obtain a unified local deformation fusion feature; then, the global axial length variation feature and the local deformation fusion feature are fused a second time, and combined with the structural characteristics of the concrete column, a global and local correlation mapping is established.

2. The method for identifying axial deformation of concrete columns based on visual images as described in claim 1, characterized in that, The process of performing continuous image sequence preprocessing, adaptively selecting local evaluation center points, and then performing adjacent frame tracking and recognition of the local evaluation center points to establish first local deformation features includes: The first image in a continuous image sequence is acquired, and concrete column texture recognition is performed on the first image to establish the texture recognition result. If the coverage of the texture recognition result meets the preset coverage threshold, then the texture recognition result is used as the main recognition result. After adaptively distributing the texture evaluation center points, the concrete columns that do not cover the texture evaluation center points are configured with grid evaluation center points. Local evaluation center points are established using the grid evaluation center point and the texture evaluation center point to perform adjacent frame tracking and recognition, and to establish the first local deformation feature.

3. The method for identifying axial deformation of concrete columns based on visual images as described in claim 2, characterized in that, After establishing the texture recognition result, the following is included: If the coverage of the texture recognition result does not meet the preset coverage threshold, a grid center point distribution instruction is generated. The grid evaluation center point configuration of the concrete column is adaptively executed according to the grid center point distribution instructions; Using the configured mesh evaluation center point as the search center, perform a texture key point association search based on the texture recognition results; After replacing the corresponding grid evaluation center point using associative search, the construction of local evaluation center points is completed.

4. The method for identifying axial deformation of concrete columns based on visual images as described in claim 3, characterized in that, The process of performing adjacent frame tracking and identification of the local evaluation center point to establish the first local deformation feature includes: Optical flow tracking of local evaluation center points is performed in adjacent frames, and tracking calibration is performed through local texture gradient field features, RGB-D depth point cloud constraints and mutual information similarity matching to perform local evaluation center point alignment. A sliding time window is established for the continuous multi-frame trajectory of each local evaluation center point, and the temporal displacement characteristics of the local evaluation center point are calculated based on the sliding time window. Using the local evaluation center point as the search zero point, feature extraction of the neighboring pixel block is performed to establish a neighborhood strain field, which includes the local strain tensor, volumetric strain, and shear strain energy density. After verifying the consistency of time-series displacement characteristics using the neighborhood strain field, a first local deformation feature is established.

5. The method for identifying axial deformation of concrete columns based on visual images as described in claim 4, characterized in that, The feature extraction of the neighboring pixel blocks includes: After setting the calibration neighborhood range, configuring the initial neighborhood region based on the search zero point and the calibration neighborhood range, performing neighborhood complexity calculation within the initial neighborhood region, and establishing the first compensation factor; Obtain the temporal displacement features corresponding to the search zero point, and establish a second compensation factor based on the temporal displacement features; After calibration of the neighborhood range using the first compensation factor and the second compensation factor, feature extraction of the neighborhood pixel block is performed.

6. The method for identifying axial deformation of concrete columns based on visual images as described in claim 1, characterized in that, The step of extracting local deformation features from the preprocessed continuous image sequence using a convolutional neural network to establish a second local deformation feature includes: Based on the continuous image sequence, local image patches containing multiple adjacent frames are extracted, and the local image patches are input into a convolutional neural network through multiple channels. The convolutional neural network is a residual encoder-decoder network based on U-Net, which outputs the deformation amplitude and confidence value of each pixel. A second local deformation feature is established based on the deformation amplitude and confidence value of each pixel.

7. The method for identifying axial deformation of concrete columns based on visual images as described in claim 1, characterized in that, The method of generating global axial length variation features using time-series identifiers of continuous image sequences and fitted axial centerlines of cylinders includes: In each frame of the image, after Canny edge detection, denoising is performed using morphological opening and closing operations to extract the outer contour of the cylinder. Multi-view geometric projection of the outer contour of the cylinder from different perspectives at the same time point is performed to establish a spatially consistent contour. Based on the spatially consistent profile, straight lines are fitted to the boundary points on both sides, and the midline of the straight line fitting results on both sides is used as the axial centerline of the cylinder. Global length variation features are extracted from frame-by-frame images based on the axial centerline of the cylinder to generate global axial length variation features.

8. The method for identifying axial deformation of concrete columns based on visual images as described in claim 1, characterized in that, The establishment of axial deformation identification results includes: The overall deformation amplitude is identified based on the axial deformation identification results, and a first deformation warning is established; The maximum deformation rate per unit time is identified based on the axial deformation identification results, and a second deformation early warning is established; Based on the first deformation warning and the second deformation warning, an abnormal deformation report is executed.

9. The method for identifying axial deformation of concrete columns based on visual images as described in claim 1, characterized in that, The establishment of axial deformation identification results also includes: The deformation monitoring cycle is reconstructed based on the axial deformation identification results; The deformation monitoring cycle of the reconstructed structure is used for the monitoring and management of concrete columns.

10. A system for recognizing axial deformation of concrete columns based on visual images, characterized in that, The system is used to implement the visual image-based axial deformation identification method for concrete columns according to any one of claims 1-9, and the system comprises: The image sequence acquisition module is used to acquire a continuous image sequence of a concrete column from different sides using at least two RGB-D cameras, wherein the interval angle between the RGB-D cameras is greater than or equal to 45 degrees. The variation feature generation module is used to perform edge detection and morphological processing on the continuous image sequence, extract the overall contour of the column, fit the axial center line of the column, and generate global axial length variation features using the time series identifier of the continuous image sequence and the fitted axial center line of the column. The first local deformation feature establishment module is used to perform continuous image sequence preprocessing, adaptively select local evaluation center points, perform adjacent frame tracking and recognition of local evaluation center points, and establish the first local deformation features. The second local deformation feature establishment module is used to extract local deformation features from the preprocessed continuous image sequence using a convolutional neural network and establish the second local deformation features. The recognition result establishment module is used to perform hierarchical fusion of the global axial length variation feature, the first local deformation feature, and the second local deformation feature to establish the axial deformation recognition result. This includes first fusing the first local deformation feature and the second local deformation feature by weighted averaging, with the weights dynamically adjusted based on the confidence values ​​of the two types of features, and integrating the deformation data of the two types of local features to obtain a unified local deformation fusion feature; then, performing a second fusion of the global axial length variation feature and the local deformation fusion feature, and establishing a global-local correlation mapping based on the structural characteristics of the concrete column.

Images