Intelligent visual identification method for construction error of large-span steel structure
By using multi-scale gradient calculation and cylindrical perspective projection model of circular tube components, combined with geometric consistency and gradient direction verification, construction errors of large-span steel structures are identified, solving the problem of pseudo-structure edge interference and achieving high-precision construction error identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA CONSTR SIXTH ENG BUREAU CIVILENG
- Filing Date
- 2026-05-08
- Publication Date
- 2026-06-05
Smart Images
Figure CN122156199A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent visual recognition technology, and in particular to an intelligent visual recognition method for construction errors in large-span steel structures. Background Technology
[0002] In architectural projects such as airport terminals, stadiums, and convention centers, large-span steel structure systems are widely used in roofs, connecting corridors, and spatial truss structures. The construction of such steel structures involves segmented hoisting, high-altitude assembly, and node connection. Due to the large number of components and complex node forms, construction errors such as member misalignment and node displacement are prone to occur. Utilizing intelligent vision methods to quickly identify these construction errors has become an important means of improving installation accuracy.
[0003] Existing intelligent visual recognition methods for steel structure construction errors mainly involve acquiring construction images of node areas and extracting geometric features of components and calculating construction error parameters based on image processing algorithms such as edge detection and contour extraction. However, in the high-altitude construction environment of large-span spatial trusses or space frames, non-structural targets such as hoisting wire ropes, safety ropes, welding cables, and temporary clamps appear as slender linear shapes in the images, highly similar to the contour edges of the members. Simultaneously, the metal surface of circular steel components easily forms high-gloss reflection bands, whose gradient abrupt changes are comparable to the response of the true contour edges. The linear shapes of non-structural targets and the high-gloss reflection bands superimpose each other, forming pseudo-structural edges in the construction images of node areas.
[0004] False structural edges have a significant negative impact on the intelligent visual recognition of construction errors in large-span steel structures. First, because the gradient response intensity of false structural edges is similar to that of true structural edges, they are difficult to eliminate using conventional threshold screening. Second, they are easily incorrectly included in the candidate set of member boundaries, leading to centerline fitting shifts and altering the geometric feature distribution of node regions, causing node positioning deviations. Furthermore, the generation of false structural edges is related to time-varying factors such as the location of non-structural targets, illumination angle, and shooting perspective, causing fluctuations in error recognition results and severely limiting the reliability of intelligent visual methods in the engineering application of error recognition in large-span steel structure construction. Summary of the Invention
[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an intelligent visual recognition method for construction errors of large-span steel structures, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent visual recognition method for construction errors in large-span steel structures, comprising: Construction images of steel structure construction node areas are collected and marked as node construction drawings. Multi-scale gradient calculations are performed on the node construction drawings to generate multi-scale gradient response maps. Candidate edge segments are extracted based on multi-scale gradient response maps, and the geometric attribute vectors of each candidate edge segment are calculated to form a set of candidate edge segments; Based on the dual edge geometric constraints of the cylindrical perspective projection of the circular tube component, a dual pairing hypothesis is constructed for the candidate edge segments in the candidate edge segment set, and a geometric consistency test is performed. Candidate edge segment pairs that pass the geometric consistency test are retained to generate a candidate structural edge pair set. Along the common normal direction of each candidate edge pair in the candidate structure edge pair set, extract the cross-edge gradient sampling sequence. For the cross-edge gradient sampling sequences corresponding to the two candidate edge segments in the same candidate edge pair, perform gradient direction opposition verification. Retain candidate edge pairs whose gradient direction opposition meets the preset threshold condition to generate an effective structure edge pair set. The midline pixel sequence of each candidate edge segment pair in the set of valid structural edge pairs is marked in the node construction drawing to generate a construction structure edge annotation map; The pre-stored design model is projected onto the same view coordinate system as the node construction drawing to generate a design reference edge map. The construction structure edge annotation map and the design reference edge map are then image registered and deviation analyzed. Based on the registration deviation, a construction error distribution map is generated, and the construction error identification results are output.
[0007] Preferably, the execution steps for generating the multi-scale gradient response map are as follows: By using image acquisition equipment deployed in the construction node area of steel structure, construction images including circular pipe components and node connection areas are acquired, and these images are marked as node construction drawings. The node construction drawings are then processed into grayscale to obtain grayscale node images. A multi-scale smoothed image sequence is obtained by Gaussian smoothing the grayscale node image with multiple preset scale parameters. The single-scale gradient magnitude map and single-scale gradient direction map corresponding to each scale are obtained by convolution operation of the smoothed image at each scale in the multi-scale smoothed image sequence using gradient operator. The single-scale gradient magnitude maps corresponding to each scale are subjected to a pixel-by-pixel maximum value operation to generate a fused gradient magnitude map; the gradient direction at the scale corresponding to the gradient magnitude of each pixel in the fused gradient magnitude map is taken as the final gradient direction of that pixel to generate a fused gradient direction map; the fused gradient magnitude map and the fused gradient direction map are merged to generate a multi-scale gradient response map.
[0008] Preferably, the execution step of forming the candidate edge segment set is as follows: By fusing gradient magnitude maps into multi-scale gradient response maps, and then processing them with non-maximum suppression and double threshold segmentation in Canny edge detection, a binary edge map is obtained. By performing edge chain tracking on the binary edge map and breaking the edge chain at locations where the gradient direction of adjacent pixels changes beyond a preset angle threshold, candidate edge segments are obtained. For each pixel position in each candidate edge segment, calculate the local orientation angle and local curvature based on the coordinates of the pixel and its preceding and following pixels; calculate the mean and variance of the gradient magnitude of all pixels in each candidate edge segment, and record the pixel length of each candidate edge segment. The principal orientation angle, average curvature, pixel length, mean gradient magnitude, and gradient magnitude variance are combined to form the geometric attribute vector corresponding to each candidate edge segment; then, each candidate edge segment and its corresponding geometric attribute vector are summarized to form a candidate edge segment set.
[0009] Preferably, the steps for constructing the dual pairing hypothesis and performing the geometric consistency test are as follows: For each candidate edge segment in the candidate edge segment set, it is sequentially taken as a hypothetical edge segment. Based on the geometric attribute vector of the hypothetical edge segment and the known outer diameter specification of the circular tube component, the predicted dual trajectory of the dual edge of each hypothetical edge segment in the pixel coordinate system of the node construction drawing is calculated through the cylindrical perspective projection model. The predicted dual trajectory includes the predicted position, predicted direction angle and predicted curvature. Centered on each predicted dual trajectory, a search band of preset width is set along the normal direction. Candidate edge segments that fall within the search band are searched in the candidate edge segment set. The candidate edge segments that are hit by the search are combined with the corresponding hypothetical edge segments to form a dual pairing hypothesis. For each dual pairing hypothesis, a geometric consistency test is performed on the candidate edge segments that are searched and the corresponding predicted dual trajectories. The positional deviation, orientation angle deviation, and curvature deviation between the candidate edge segments that are searched and the predicted dual trajectories are calculated. Dual pairing hypotheses in which the positional deviation, orientation angle deviation, and curvature deviation all fall within the corresponding allowable range are marked as passing the geometric consistency test. The hypothetical edge segments in the dual pairing hypothesis that have passed the geometric consistency test and the candidate edge segments that have been found in the search are taken as candidate edge segment pairs. All candidate edge segment pairs are summarized to generate a set of candidate structural edge pairs.
[0010] Preferably, the step of performing gradient direction opposition verification is as follows: Along the common normal direction of each candidate edge segment pair in the candidate structure edge pair set, K sampling positions are set at equal intervals; At each sampling position, a sampling window with a half width of w pixels is taken along the common normal direction, centered on the corresponding edge pixel on the two candidate edge segments. Within the sampling window, gray values are extracted pixel by pixel along the common normal direction and first-order difference operation is performed to generate cross-edge gradient sampling sequences corresponding to the two candidate edge segments. Perform polarity inversion matching on the two cross-edge gradient sampling sequences at each sampling location, and calculate the local opposition score at each sampling location. The mean of the local opposition scores at all K sampling locations is used to obtain the gradient direction opposition R of the candidate edge segment pair; The gradient direction opposition degree R is compared with a preset threshold condition, and candidate edge segment pairs with gradient direction opposition degree R not lower than the preset threshold are retained to generate a set of effective structural edge pairs.
[0011] Preferably, the execution steps for generating the construction structure edge annotation diagram are as follows: For each candidate edge pair in the set of effective structural edge pairs, calculate the midpoint pixel coordinates between the two candidate edge pairs pixel by pixel along the common normal direction to generate the midline pixel sequence; For each midline pixel sequence, the positions of discontinuous pixels are filled in by pixel connectivity interpolation to generate a continuous midline pixel chain; All continuous centerline pixel chains are labeled in the pixel coordinate system of the node construction drawing to generate the edge labeling map of the construction structure.
[0012] Preferably, the execution step of generating a construction error distribution map based on the registration deviation and outputting the construction error identification result is as follows: The pre-stored design model is obtained, and based on the shooting parameters of the node construction drawing, the center lines of each member and the node position in the design model are projected onto the pixel coordinate system of the node construction drawing through perspective projection transformation to generate the design reference edge map. Register each continuous centerline pixel chain in the construction structure edge annotation drawing with the projection centerline of the corresponding member in the design reference edge drawing, calculate the pixel-by-pixel offset of each corresponding member in the common normal direction, and generate the pixel distribution of offset of each member. Based on the pixel coordinate difference between the extended intersection pixel positions of multiple continuous centerline pixel chains in the construction structure edge annotation map and the corresponding node projected pixel positions in the design reference edge map, the node-by-node deviation pixel amount is calculated. The pixel distribution of each member offset and the pixel quantity of each node deviation are mapped to the pixel coordinate system of the node construction image to generate a construction error distribution map. Based on the shooting parameters, each offset pixel quantity and deviation pixel quantity is converted into the actual physical size, and the construction error identification result is output.
[0013] As described above, the intelligent visual recognition method for large-span steel structure construction errors provided by this invention has at least the following beneficial effects: The intelligent visual recognition method for large-span steel structure construction errors provided by this invention constructs candidate edge segments in the construction node area image. Based on the dual pairing assumption of the cylindrical perspective projection model of the circular tube component, it actively generates predicted dual trajectories and performs geometric consistency checks using the dual edge prediction mapping method. This transforms the traditional exhaustive pairing screening paradigm into an active prediction verification paradigm based on a physical model. As a result, pseudo-structure edges are naturally excluded in the pairing stage because they do not have the physical basis of the cylindrical projection of the circular tube component. There is no need to build a separate pseudo-structure edge recognition model. This fundamentally solves the technical problem that it is difficult to distinguish between pseudo-structure edges and real structure edges in construction images at the single edge level, and improves the accuracy of construction error recognition.
[0014] Meanwhile, this invention performs gradient direction opposition verification on candidate structural edge pairs. Utilizing the physical and optical properties that the gray-scale gradient transition morphology at the dual edges of circular tube components has a polar correspondence, it performs secondary verification of the authenticity of candidate edge pairs from the gradient transition morphology level. This, together with the geometric consistency check, forms a dual filtering mechanism from geometric position constraints to optical and physical constraints, effectively eliminating mispairing caused by accidental mismatches due to different non-structural targets. This further improves the reliability of effective structural edge recognition. Furthermore, by projecting the design model into a design reference edge map, the registration and deviation analysis of construction errors are completed in the image domain. This ensures that the entire process is completed at the image data level, guaranteeing the stability and consistency of the construction error recognition results. Attached Figure Description
[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.
[0016] Figure 1 This is a flowchart illustrating the intelligent visual recognition method for large-span steel structure construction errors according to the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Please see Figure 1 As shown, this invention provides an intelligent visual recognition method for construction errors in large-span steel structures, comprising the following steps: S1. Collect construction images of the steel structure construction node area, mark them as node construction drawings, perform multi-scale gradient calculations on the node construction drawings, and generate multi-scale gradient response maps. In this embodiment, it should be specifically explained that the execution steps for generating the multi-scale gradient response map are as follows: By using image acquisition equipment deployed in the construction node area of steel structure, construction images including circular pipe components and node connection areas are acquired, and these images are marked as node construction drawings. The node construction drawings are then processed into grayscale to obtain grayscale node images. It should be noted that image acquisition equipment includes, but is not limited to, industrial cameras, drone-mounted cameras, or portable high-resolution cameras. The placement and shooting angle of the image acquisition equipment are determined based on the spatial location of the construction node area.
[0019] A multi-scale smoothed image sequence is obtained by Gaussian smoothing the grayscale node image with multiple preset scale parameters. The single-scale gradient magnitude map and single-scale gradient direction map corresponding to each scale are obtained by convolution operation of the smoothed image at each scale in the multi-scale smoothed image sequence using gradient operator. It should be noted that the selection of dimensional parameters is related to the diameter range of the circular pipe components at the steel structure construction site and the shooting distance.
[0020] In one specific embodiment, the diameter of the circular tube component is commonly in the range of 60mm to 300mm, the shooting distance is 3m to 15m, and the corresponding Gaussian kernel standard deviation is set to 1 pixel, 2 pixels, 4 pixels and 8 pixels, for a total of 4 scales. A smaller Gaussian kernel standard deviation is beneficial to retain the edge details of the thin-diameter tube component, while a larger Gaussian kernel standard deviation is beneficial to enhance the gradient continuity of the contour edge of the thick-diameter tube component.
[0021] The single-scale gradient magnitude maps corresponding to each scale are subjected to a pixel-by-pixel maximum value operation to generate a fused gradient magnitude map; the gradient direction at the scale corresponding to the gradient magnitude of each pixel in the fused gradient magnitude map is taken as the final gradient direction of that pixel to generate a fused gradient direction map; the fused gradient magnitude map and the fused gradient direction map are merged to generate a multi-scale gradient response map.
[0022] It should be noted that the pixel-wise maximum value operation refers to taking the maximum gradient magnitude at each scale for the same pixel location, so that the fused gradient magnitude map retains the edge response information of components with different pipe diameters. The gradient direction of each pixel in the fused gradient direction map is taken from the scale with the largest gradient magnitude. This scale corresponds to the detection scale with the strongest edge response at that pixel, and its gradient direction best reflects the true orientation of the edge at that pixel. The multi-scale gradient response map includes the fused gradient magnitude map and the fused gradient direction map, providing gradient magnitude and gradient direction information for the subsequent candidate edge segment extraction in S2.
[0023] It should be noted that grayscale conversion, Gaussian smoothing, and gradient operator convolution are all well-known techniques in the field of image processing, and will not be described in detail in this invention. Optional gradient operators include, but are not limited to, the Sobel operator, the Scharr operator, or the Prewitt operator.
[0024] S2. Extract candidate edge segments based on the multi-scale gradient response map, and calculate the geometric attribute vector of each candidate edge segment to form a set of candidate edge segments; In this embodiment, it should be specifically explained that the execution steps for forming the candidate edge segment set are as follows: By fusing gradient magnitude maps into multi-scale gradient response maps, and then processing them with non-maximum suppression and double threshold segmentation in Canny edge detection, a binary edge map is obtained. By performing edge chain tracking on the binary edge map and breaking the edge chain at locations where the gradient direction of adjacent pixels changes beyond a preset angle threshold, candidate edge segments are obtained. It should be noted that the preset angle threshold is used to control the consistency of gradient direction within candidate edge segments. In a specific embodiment, the preset angle threshold is set to 30 degrees. That is, when the gradient direction difference between two adjacent pixels in the edge chain exceeds 30 degrees, the edge chain is broken at that position, so that each candidate edge segment maintains local consistency of gradient direction within its extension range.
[0025] For each pixel position in each candidate edge segment, calculate the local orientation angle and local curvature based on the coordinates of the pixel and its preceding and following pixels; calculate the mean and variance of the gradient magnitude of all pixels in each candidate edge segment, and record the pixel length of each candidate edge segment. The principal orientation angle, average curvature, pixel length, mean gradient magnitude, and gradient magnitude variance are combined to form the geometric attribute vector corresponding to each candidate edge segment; then, each candidate edge segment and its corresponding geometric attribute vector are summarized to form a candidate edge segment set.
[0026] It should be noted that the local orientation angle refers to the instantaneous extension direction of the edge at a certain pixel in the candidate edge segment, which is calculated by the arctangent of the coordinate difference between the pixels before and after that pixel. Local curvature refers to the degree of curvature of the edge at a pixel, which is calculated from the curvature circle formed by the pixel and its three adjacent pixels.
[0027] The principal orientation angle is the arithmetic mean of the local orientation angles of all pixels; The average curvature is the arithmetic mean of the local curvatures of all pixels; The pixel length is the total number of pixels contained in the candidate edge segment; The gradient magnitude mean and gradient magnitude variance are the arithmetic mean and statistical variance of the gradient magnitudes for all pixels, respectively. These five components describe the characteristics of the candidate edge segment from five dimensions: direction, curvature, length, gradient intensity, and gradient uniformity, providing a basis for comparison in the geometric consistency test of the dual pairing hypothesis in S3. It should be noted that nonmaximum suppression and double threshold segmentation are both well-known edge detection techniques in the field of image processing, and edge chain tracing is a well-known connected component analysis technique in the field of image processing. These will not be elaborated upon in this invention.
[0028] S3. Based on the dual edge geometric constraints of the cylindrical perspective projection of the circular tube component, construct dual pairing assumptions for the candidate edge segments in the candidate edge segment set, and perform geometric consistency test. Retain the candidate edge segment pairs that pass the geometric consistency test to generate a candidate structural edge pair set. In this embodiment, it is necessary to specifically explain the geometric constraints of the dual edges in the perspective projection of the cylindrical surface of the circular tube component: the circular tube component is a cylindrical structure with a circular cross-section. Under perspective imaging conditions, the cylindrical surface of the circular tube component generates two visible contour edges on the image plane, corresponding to the tangent projection positions on both sides of the circular tube component. These two contour edges constitute a set of dual edges. Since the circular tube component has a definite outer diameter and spatial orientation, the relative position, extension direction, and curvature change of its dual edges in the image are strictly constrained by the cylindrical perspective projection model. Although pseudo-structural edges may resemble real structural edges at the level of a single edge, they do not have the physical basis to form the above-mentioned dual geometric relationship with the other edge. Based on dual edge geometric constraints, this invention employs a dual edge prediction mapping method to construct dual pairing hypotheses and perform geometric consistency checks. The execution steps are as follows: For each candidate edge segment in the candidate edge segment set, it is sequentially taken as a hypothetical edge segment. Based on the geometric attribute vector of the hypothetical edge segment and the known outer diameter specification of the circular tube component, the predicted dual trajectory of the dual edge of each hypothetical edge segment in the pixel coordinate system of the node construction drawing is calculated through the cylindrical perspective projection model. The predicted dual trajectory includes the predicted position, predicted direction angle and predicted curvature. It should be noted that predicting the dual trajectory refers to calculating the spatial trajectory that the dual edge should appear in the node construction drawing based on the assumed edge segment and the cylindrical perspective projection model.
[0029] The formulas for calculating the predicted position, predicted orientation angle, and predicted curvature are as follows: If we assume the principal direction angle of the edge segment is... Assuming the average curvature of the edge segment is ks, and assuming the coordinates of the i-th pixel in the edge segment are... The outer diameter of the circular tube component is D, the camera focal length is f, the pixel size is δ, and the shooting distance is Z; The formula for calculating the predicted normal spacing d is: ; Predict the predicted position of the i-th pixel in the dual trajectory The calculation formula is: , ; The predicted direction angle is equal to the predominant direction angle of the assumed edge segment. The predicted curvature is equal to the average curvature ks of the assumed edge segment. Both are based on the physical constraint that the extension direction and bending trend of the dual edge of the same circular tube component should be consistent in the node construction drawing.
[0030] When multiple pipe diameters exist at the construction site, multiple predicted dual trajectories can be generated for the same assumed edge segment, each corresponding to a different pipe diameter. In a specific embodiment, the outer diameter of the circular pipe component is 60mm, 114mm, 159mm, 219mm, and 299mm. The camera focal length is 25mm, the pixel size is 3.45μm, and the shooting distance ranges from 3m to 15m.
[0031] Centered on each predicted dual trajectory, a search band of preset width is set along the normal direction. Candidate edge segments that fall within the search band are searched in the candidate edge segment set. The candidate edge segments that are hit by the search are combined with the corresponding hypothetical edge segments to form a dual pairing hypothesis. It should be noted that the preset width of the search band is used to accommodate construction and installation deviations and image quantization errors. In a specific embodiment, the search band width is set to ±15% of the predicted normal spacing d. By performing directional search through predicted dual trajectories, the search range is limited to a physically reasonable area compared to exhaustively combining candidate edge segments.
[0032] For each dual pairing hypothesis, a geometric consistency test is performed on the candidate edge segments that are searched and the corresponding predicted dual trajectories. The positional deviation, orientation angle deviation, and curvature deviation between the candidate edge segments that are searched and the predicted dual trajectories are calculated. Dual pairing hypotheses in which the positional deviation, orientation angle deviation, and curvature deviation all fall within the corresponding allowable range are marked as passing the geometric consistency test. It should be noted that the positional deviation is the average normal distance from each pixel of the candidate edge segment hit by the search to the corresponding position of the predicted dual trajectory; The orientation angle deviation is the absolute value of the difference between the principal orientation angle of the candidate edge segment hit by the search and the predicted orientation angle of the predicted dual trajectory; Curvature bias is the absolute value of the difference between the average curvature of the candidate edge segments hit in the search and the predicted curvature of the predicted dual trajectory.
[0033] The allowable ranges include the positional deviation allowable range, the orientation angle deviation allowable range, and the curvature deviation allowable range. In one specific embodiment, the positional deviation allowable range is 0 to 3 pixels, the orientation angle deviation allowable range is 0 to 5 degrees, and the curvature deviation allowable range is 0 to 0.005 pixels.
[0034] The hypothetical edge segments in the dual pairing hypothesis that have passed the geometric consistency test and the candidate edge segments that have been found in the search are taken as candidate edge segment pairs. All candidate edge segment pairs are summarized to generate a set of candidate structural edge pairs.
[0035] It should be noted that this step employs a dual edge prediction mapping method. Using each candidate edge segment in the candidate edge segment set as a hypothetical edge segment, the method actively predicts the position, orientation, and curvature of its dual edge in the coordinate system of the node construction drawing elements based on the cylindrical perspective projection model of the circular tube component. This generates a predicted dual trajectory and performs a geometric consistency check within a limited search zone, transforming the traditional exhaustive pairing screening paradigm into a physical model-based active prediction verification paradigm. Through this method, the basis for pairing verification is elevated from statistical feature comparison between edge segments to geometric prediction verification based on the cylindrical perspective projection physical model. Pseudo-structural edges, lacking the physical basis of cylindrical projection for circular tube components, cannot be matched by the predicted dual trajectory and cannot generate a valid predicted dual trajectory; therefore, they are naturally excluded during the pairing stage, eliminating the need for a separate pseudo-structural edge identification model.
[0036] S4. Extract the cross-edge gradient sampling sequence along the common normal direction of each candidate edge segment pair in the candidate structure edge pair set. For the cross-edge gradient sampling sequences corresponding to the two candidate edge segments in the same candidate edge segment pair, perform gradient direction opposition verification. Retain candidate edge segment pairs whose gradient direction opposition meets the preset threshold condition to generate an effective structure edge pair set. In this embodiment, it should be specifically explained that the execution steps for generating the set of valid structural edge pairs are as follows: In this embodiment, it is necessary to specifically explain the physical basis for the gradient direction opposition verification: the circular tube component appears as a cylindrical projection area of a certain width in the image. At one edge of this area, the gray level transitions from the background to the component surface, with a positive gradient jump; at the other edge, the gray level transitions from the component surface to the background, with a negative gradient jump. The gray level gradient changes at the two edges are not only opposite in direction, but their gradient transition patterns also have a corresponding relationship—under the same material and lighting conditions, the gradient transition curve at one edge, after polarity reversal, should highly match the gradient transition curve at the other edge. The candidate edge pairs retained through the geometric consistency check in S3 have satisfied the geometric position constraints of the dual edges, but the geometric constraints cannot exclude the mispairing of two pseudo-structural edges that just meet the spacing and direction conditions but come from different non-structural targets. The gradient direction opposition verification verifies the polarity correspondence of the gray level gradient transition patterns at the edges, thus performing a secondary verification of the candidate edge pairs from an optical physics perspective.
[0037] Based on the above physical principles, the steps for performing gradient direction opposition verification are as follows: Along the common normal direction of each candidate edge segment pair in the candidate structure edge pair set, K sampling positions are set at equal intervals; It should be noted that the common normal direction refers to the direction perpendicular to the common extension direction of the two candidate edge segments. Each sampling position is uniformly distributed along the common extension direction of the two candidate edge segments. In this embodiment, K is set to 10. The purpose of using multiple sampling positions instead of a single position is to eliminate the interference of local noise, occlusion, or coating inhomogeneity on the single-point gradient.
[0038] At each sampling position, a sampling window with a half width of w pixels is taken along the common normal direction, centered on the corresponding edge pixel on the two candidate edge segments. Within the sampling window, gray values are extracted pixel by pixel along the common normal direction and first-order difference operation is performed to generate cross-edge gradient sampling sequences corresponding to the two candidate edge segments. It should be noted that the cross-edge gradient sampling sequence is a one-dimensional numerical sequence formed by extracting grayscale gradient values pixel by pixel along the common normal direction with the edge pixels as the center. Each element in the sequence corresponds to the grayscale gradient value at a pixel position, and the sequence length is 2w+1. The value of the half-width w should cover the complete gradient change of the edge transition region.
[0039] In this embodiment, w is set to 5 pixels, that is, the total width of the sampling window is 11 pixels, and the corresponding sequence length is 11.
[0040] Let the cross-edge gradient sampling sequence corresponding to the first candidate edge segment at the k-th sampling position be... The cross-edge gradient sampling sequence corresponding to the second candidate edge segment is: Where t = -w, -w+1, ..., 0, ..., w-1, w, t = 0 corresponds to the position of the edge pixel. and Each element is the grayscale gradient value at the corresponding position.
[0041] Perform polarity inversion matching on the two cross-edge gradient sampling sequences at each sampling location, and calculate the local opposition score at each sampling location. It should be noted that the cross-edge gradient sampling sequence corresponding to the second candidate edge segment... Take the opposite of each element in the equation, and you will get... Then calculate and The normalized cross-correlation value between them is used as the local opposition score at the k-th sampling location. The calculation formula is as follows: ; It should be noted that the numerator of the formula is the sum of element-wise multiplications of the two sequences, and the denominator is the product of the moduli of the two sequences. The result is a dimensionless value between -1 and 1. When the value approaches 1, it indicates that the two sequences have a high degree of morphological similarity after polarity reversal, meaning that the gradient transitions at the two edges have a strong polarity correspondence; when When the value approaches 0 or is negative, it indicates that the two do not have a polarity correspondence.
[0042] The mean of the local opposition scores at all K sampling locations is used to obtain the gradient direction opposition R of the candidate edge segment pair; The formula for calculating the gradient direction oppositeness R is: ; The gradient direction opposition degree R is compared with a preset threshold condition, and candidate edge segment pairs with gradient direction opposition degree R not lower than the preset threshold are retained to generate a set of effective structural edge pairs.
[0043] It should be noted that, in one specific embodiment, the preset threshold is set to 0.7. Based on actual construction image testing, the gradient direction contrast of the true paired edges of circular pipe components is typically between 0.8 and 0.98, while the gradient direction contrast of mispairings accidentally formed by different non-structural targets is typically below 0.5.
[0044] A gradient transition envelope matching method is employed to perform secondary verification on the candidate structural edge pair set output by S3. Utilizing the physical-optical property that the gray-scale gradient transition morphology at the dual edges of the circular tube component exhibits a polarity correspondence, the authenticity of candidate edge pairs is verified from the perspective of gradient transition morphology by extracting cross-edge gradient sampling sequences and performing polarity inversion matching. This method leverages the complete gradient distribution information within the edge transition region, ensuring that even if two pseudo-structural edges have exactly opposite gradient directions at a single point, they will be excluded because their gradient transition envelope morphology lacks a polarity correspondence. This, combined with the geometric constraint verification of S3, forms a dual filtering mechanism from geometric location to optical physics, further improving the reliability of the effective structural edge pair set.
[0045] S5. Mark the midline pixel sequence of each candidate edge segment pair in the set of valid structural edge pairs in the node construction drawing to generate a construction structure edge annotation map; In this embodiment, it should be specifically explained that the execution steps for generating the construction structure edge annotation diagram are as follows: For each candidate edge pair in the set of effective structural edge pairs, calculate the midpoint pixel coordinates between the two candidate edge pairs pixel by pixel along the common normal direction to generate the midline pixel sequence; It should be noted that the midpoint pixel coordinates are obtained by the arithmetic mean of the corresponding pixel coordinates of the two candidate edge segments in the same common normal direction. The midline pixel sequence is an ordered set of pixels formed by arranging all midpoint pixel coordinates in the extension direction. This midline pixel sequence represents the projection position of the centerline of the corresponding circular pipe component in the node construction drawing.
[0046] For each midline pixel sequence, the positions of discontinuous pixels are filled in by pixel connectivity interpolation to generate a continuous midline pixel chain; It should be noted that pixel connectivity interpolation is a well-known technique in the field of image processing, and will not be described in detail in this invention; All continuous centerline pixel chains are labeled in the pixel coordinate system of the node construction drawing to generate the edge labeling map of the construction structure.
[0047] It should be noted that the construction structure edge annotation diagram and the node construction diagram have the same pixel coordinate system.
[0048] S6. Project the pre-stored design model onto the same view coordinate system as the node construction drawing to generate the design reference edge map. Perform image registration and deviation analysis on the construction structure edge annotation map and the design reference edge map. Generate the construction error distribution map based on the registration deviation and output the construction error identification results.
[0049] In this embodiment, it should be specifically explained that the execution steps of generating a construction error distribution map based on the registration deviation and outputting the construction error identification result are as follows: The pre-stored design model is obtained, and based on the shooting parameters of the node construction drawing, the center lines of each member and the node position in the design model are projected onto the pixel coordinate system of the node construction drawing through perspective projection transformation to generate the design reference edge map. It should be noted that the design model is a three-dimensional digital model established based on the design drawings before the steel structure construction, containing the spatial coordinates of the centerlines of each member and the spatial coordinates of each node. The shooting parameters include camera intrinsic and extrinsic parameters; the design baseline edge map and the node construction drawing have the same pixel coordinate system, and each pixel chain in the map corresponds to the theoretical projection position of the centerline of each member and the node position under the design state.
[0050] Perspective projection transformation is a well-known technique in the field of computer vision, and will not be described in detail in this invention.
[0051] Register each continuous centerline pixel chain in the construction structure edge annotation drawing with the projection centerline of the corresponding member in the design reference edge drawing, calculate the pixel-by-pixel offset of each corresponding member in the common normal direction, and generate the pixel distribution of offset of each member. It should be noted that member-by-member registration involves calculating the angular difference and positional distance between each continuous centerline pixel chain in the construction structure edge annotation drawing and each projected centerline in the design reference edge drawing, establishing a correspondence between the members with the smallest angular difference and positional distance. The calculation method for the pixel-by-pixel offset is as follows: for each member with an established correspondence, a perpendicular line is drawn from the common normal direction along each pixel position in the continuous centerline pixel chain. The intersection of this perpendicular line with the corresponding projected centerline is found, and the pixel distance from that pixel to the intersection point is the offset pixel amount at that position. The offset pixel amounts of all pixel positions in the continuous centerline pixel chain are arranged sequentially to form the member-by-member offset pixel distribution, reflecting the degree of deviation between the construction state and the design state at each position along the length of the member.
[0052] Based on the pixel coordinate difference between the extended intersection pixel positions of multiple continuous centerline pixel chains in the construction structure edge annotation map and the corresponding node projected pixel positions in the design reference edge map, the node-by-node deviation pixel amount is calculated. It should be noted that the extended intersection pixel position is obtained by extending and intersecting the continuous centerline pixel chain of adjacent members in the edge annotation diagram of the construction structure. This intersection position corresponds to the image projection position of the node center under the actual construction state. The method for calculating the node-by-node deviation pixel amount is as follows: calculate the Euclidean pixel distance between the extended intersection pixel position and the corresponding node projection pixel position in the design reference edge diagram. The deviation pixel amount reflects the comprehensive deviation between the construction state and the design state of the node on the image plane.
[0053] The pixel distribution of each member offset and the pixel quantity of each node deviation are mapped to the pixel coordinate system of the node construction image to generate a construction error distribution map. Based on the shooting parameters, each offset pixel quantity and deviation pixel quantity is converted into the actual physical size, and the construction error identification result is output.
[0054] It should be noted that the formula for converting pixel count to actual physical size is as follows: Let the offset pixel amount be Given a camera focal length of f, a pixel size of δ, and a shooting distance of Z, the corresponding actual physical offset L is: .
[0055] In this embodiment, it should be specifically noted that the above formulas are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0056] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. An intelligent visual recognition method for construction errors in large-span steel structures, characterized in that, include: Construction images of steel structure construction node areas are collected and marked as node construction drawings. Multi-scale gradient calculations are performed on the node construction drawings to generate multi-scale gradient response maps. Candidate edge segments are extracted based on multi-scale gradient response maps, and the geometric attribute vectors of each candidate edge segment are calculated to form a set of candidate edge segments; Based on the dual edge geometric constraints of the cylindrical perspective projection of the circular tube component, a dual pairing hypothesis is constructed for the candidate edge segments in the candidate edge segment set, and a geometric consistency test is performed. Candidate edge segment pairs that pass the geometric consistency test are retained to generate a candidate structural edge pair set. Along the common normal direction of each candidate edge pair in the candidate structure edge pair set, extract the cross-edge gradient sampling sequence. For the cross-edge gradient sampling sequences corresponding to the two candidate edge segments in the same candidate edge pair, perform gradient direction opposition verification. Retain candidate edge pairs whose gradient direction opposition meets the preset threshold condition to generate an effective structure edge pair set. The midline pixel sequence of each candidate edge segment pair in the set of valid structural edge pairs is marked in the node construction drawing to generate a construction structure edge annotation map; The pre-stored design model is projected onto the same view coordinate system as the node construction drawing to generate a design reference edge map. The construction structure edge annotation map and the design reference edge map are then image registered and deviation analyzed. Based on the registration deviation, a construction error distribution map is generated, and the construction error identification results are output.
2. The intelligent visual recognition method for construction errors of large-span steel structures according to claim 1, characterized in that: The steps for generating the multi-scale gradient response map are as follows: By using image acquisition equipment deployed in the construction node area of steel structure, construction images including circular pipe components and node connection areas are acquired, and these images are marked as node construction drawings. The node construction drawings are then processed into grayscale to obtain grayscale node images. A multi-scale smoothed image sequence is obtained by Gaussian smoothing the grayscale node image with multiple preset scale parameters. The single-scale gradient magnitude map and single-scale gradient direction map corresponding to each scale are obtained by convolution operation of the smoothed image at each scale in the multi-scale smoothed image sequence using gradient operator. The single-scale gradient magnitude maps corresponding to each scale are subjected to a pixel-by-pixel maximum value operation to generate a fused gradient magnitude map; the gradient direction at the scale corresponding to the gradient magnitude of each pixel in the fused gradient magnitude map is taken as the final gradient direction of that pixel to generate a fused gradient direction map. The fused gradient magnitude map and the fused gradient direction map are merged to generate a multi-scale gradient response map.
3. The intelligent visual recognition method for construction errors of large-span steel structures according to claim 1, characterized in that: The steps for forming the candidate edge segment set are as follows: By fusing gradient magnitude maps into multi-scale gradient response maps, and then processing them with non-maximum suppression and double threshold segmentation in Canny edge detection, a binary edge map is obtained. By performing edge chain tracking on the binary edge map and breaking the edge chain at locations where the gradient direction of adjacent pixels changes beyond a preset angle threshold, candidate edge segments are obtained. For each pixel position in each candidate edge segment, calculate the local orientation angle and local curvature based on the coordinates of the pixel and its preceding and following pixels; calculate the mean and variance of the gradient magnitude of all pixels in each candidate edge segment, and record the pixel length of each candidate edge segment. The principal orientation angle, average curvature, pixel length, mean gradient magnitude, and gradient magnitude variance are combined to form the geometric attribute vector corresponding to each candidate edge segment; then, each candidate edge segment and its corresponding geometric attribute vector are summarized to form a candidate edge segment set.
4. The intelligent visual recognition method for construction errors of large-span steel structures according to claim 1, characterized in that: The steps for constructing the dual pairing hypothesis and performing the geometric consistency test are as follows: For each candidate edge segment in the candidate edge segment set, it is sequentially taken as a hypothetical edge segment. Based on the geometric attribute vector of the hypothetical edge segment and the known outer diameter specification of the circular tube component, the predicted dual trajectory of the dual edge of each hypothetical edge segment in the pixel coordinate system of the node construction drawing is calculated through the cylindrical perspective projection model. The predicted dual trajectory includes the predicted position, predicted direction angle and predicted curvature. Centered on each predicted dual trajectory, a search band of preset width is set along the normal direction. Candidate edge segments that fall within the search band are searched in the candidate edge segment set. The candidate edge segments that are hit by the search are combined with the corresponding hypothetical edge segments to form a dual pairing hypothesis. For each dual pairing hypothesis, a geometric consistency test is performed on the candidate edge segments that are searched and the corresponding predicted dual trajectories. The positional deviation, orientation angle deviation, and curvature deviation between the candidate edge segments that are searched and the predicted dual trajectories are calculated. Dual pairing hypotheses in which the positional deviation, orientation angle deviation, and curvature deviation all fall within the corresponding allowable range are marked as passing the geometric consistency test. The hypothetical edge segments in the dual pairing hypothesis that have passed the geometric consistency test and the candidate edge segments that have been found in the search are taken as candidate edge segment pairs. All candidate edge segment pairs are summarized to generate a set of candidate structural edge pairs.
5. The intelligent visual recognition method for construction errors of large-span steel structures according to claim 1, characterized in that: The steps for performing gradient direction opposition verification are as follows: Along the common normal direction of each candidate edge segment pair in the candidate structure edge pair set, K sampling positions are set at equal intervals; At each sampling position, a sampling window with a half width of w pixels is taken along the common normal direction, centered on the corresponding edge pixel on the two candidate edge segments. Within the sampling window, gray values are extracted pixel by pixel along the common normal direction and first-order difference operation is performed to generate cross-edge gradient sampling sequences corresponding to the two candidate edge segments. Perform polarity inversion matching on the two cross-edge gradient sampling sequences at each sampling location, and calculate the local opposition score at each sampling location. The mean of the local opposition scores at all K sampling locations is used to obtain the gradient direction opposition R of the candidate edge segment pair; The gradient direction opposition degree R is compared with a preset threshold condition, and candidate edge segment pairs with gradient direction opposition degree R not lower than the preset threshold are retained to generate a set of effective structural edge pairs.
6. The intelligent visual recognition method for construction errors of large-span steel structures according to claim 1, characterized in that: The steps for generating the construction structure edge annotation diagram are as follows: For each candidate edge pair in the set of effective structural edge pairs, calculate the midpoint pixel coordinates between the two candidate edge pairs pixel by pixel along the common normal direction to generate the midline pixel sequence; For each midline pixel sequence, the positions of discontinuous pixels are filled in by pixel connectivity interpolation to generate a continuous midline pixel chain; All continuous centerline pixel chains are labeled in the pixel coordinate system of the node construction drawing to generate the edge labeling map of the construction structure.
7. The intelligent visual recognition method for construction errors of large-span steel structures according to claim 1, characterized in that: The steps for generating a construction error distribution map based on the registration deviation and outputting the construction error identification results are as follows: The pre-stored design model is obtained, and based on the shooting parameters of the node construction drawing, the center lines of each member and the node position in the design model are projected onto the pixel coordinate system of the node construction drawing through perspective projection transformation to generate the design reference edge map. Register each continuous centerline pixel chain in the construction structure edge annotation drawing with the projection centerline of the corresponding member in the design reference edge drawing, calculate the pixel-by-pixel offset of each corresponding member in the common normal direction, and generate the pixel distribution of offset of each member. Based on the pixel coordinate difference between the extended intersection pixel positions of multiple continuous centerline pixel chains in the construction structure edge annotation map and the corresponding node projected pixel positions in the design reference edge map, the node-by-node deviation pixel amount is calculated. The pixel distribution of each member offset and the pixel quantity of each node deviation are mapped to the pixel coordinate system of the node construction image to generate a construction error distribution map. Based on the shooting parameters, each offset pixel quantity and deviation pixel quantity is converted into the actual physical size, and the construction error identification result is output.