A texture reconstruction method for steel structure circular pipe intersecting welding joint
By using a sparse/semi-dense information reconstruction method, the problems of high computational cost and poor stability in the texture reconstruction of intersecting weld nodes of circular pipes in steel structures are solved, achieving efficient and robust texture reconstruction, which is suitable for rapid modeling and defect display in engineering sites.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA RAILWAY CONSTR CORP LTD
- Filing Date
- 2026-04-27
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies suffer from high computational costs and poor stability in texture reconstruction of intersecting welded joints of circular tubes in steel structures, making it difficult to meet the demand for "3D models with realistic textures" in engineering sites. In particular, holes and noise are prone to occur during multi-view matching, and there is a lack of adaptation and constraints for the parameterization of circular tubes and the Boolean geometry features of intersections.
A sparse/semi-dense information reconstruction method is adopted. Sparse point clouds are generated by multi-angle image acquisition, parametric geometric modeling and texture mapping are performed, and semi-dense point clouds are generated by combining adaptive clustering and least squares optimization. Visual filtering is then applied to achieve texture reconstruction.
It significantly reduces computational and storage overhead, improves the efficiency and robustness of texture reconstruction, and is suitable for engineering inspection, BIM and digital twin systems, facilitating rapid on-site modeling and preservation of defect records.
Smart Images
Figure CN122492907A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of maintenance technology for large-span spatial steel structures, and in particular to a texture reconstruction method for intersecting welded joints of circular tubes in steel structures. Background Technology
[0002] Large-span spatial steel structures (such as space frames, trusses, and domes) are widely used in stadiums, convention centers, transportation hubs, and industrial plants. These structures feature numerous and diverse circular tube intersecting weld nodes (T-shaped, K-shaped, N-shaped, etc.) with large spans. The weld toes / welds are small in size, and the surface metal is highly reflective. Geometric models alone are insufficient to meet the needs of visualizing defects, preserving maintenance records, conducting detailed acceptance inspections, and showcasing operational and maintenance processes. The demand for "3D models with realistic textures" is increasingly urgent on-site. Texture mapping, while retaining the advantages of geometric simplification and parametric mapping, can intuitively present surface information such as rust, welding spatter, grinding marks, and coating blistering / peeling. It is a crucial step from "geometric reconstruction" to "visualized diagnosis." However, deep cavities and strong occlusions exist at the intersection points, and the repeated textures on the cylindrical surfaces combined with the high reflectivity of the metal make multi-view matching prone to degradation and generate holes and noise. Traditional processes that rely on dense reconstruction are therefore computationally expensive and unstable in this area, and multi-source splicing is prone to seams and color differences. Therefore, there is an urgent need for a texture mapping method that is both efficient and robust for intersecting weld nodes.
[0003] Currently, there are numerous patents for algorithms related to the three-dimensional reconstruction of space frame structures.
[0004] For example, the patent with application number 202410603708.X proposed a method, device, computer equipment and storage medium for three-dimensional reconstruction of space frame structures. Its feature is that it establishes the correspondence between two-dimensional and three-dimensional structures by identifying the boundaries of space frame members in the image, determining the four corner nodes, and matching the original image with the two-dimensional mapping, and finally obtains the space frame structure model.
[0005] The patent application with application number 202410142534.1 proposes a method and device for three-dimensional reconstruction of a grid structure based on image processing and deep learning. It uses a deep learning model to obtain the position and set information of the sphere nodes, and then judges the connection status of each sphere node to realize the reconstruction of the structural model.
[0006] The patent application number 201710045544.3 proposes a method for reverse modeling of a ball-linked steel structure space frame. It uses three-dimensional laser scanning to obtain a dense point cloud of the space frame, and then completes the reverse three-dimensional modeling of the space frame structure based on point cloud segmentation.
[0007] The patent application with application number 201710044770.X proposes a method for reverse modeling of tubular welded steel structure space frame. It uses three-dimensional laser scanning to obtain dense point cloud of tubular welded steel structure space frame, and then constructs the spatial topological relationship between each tubular component based on cylindrical surface fitting, and finally completes the reverse three-dimensional modeling of the structure.
[0008] The aforementioned solutions primarily focus on geometric / topological reconstruction, which can be used for shape restoration, component size acquisition, and assembly relationship recovery. However, they generally fail to address the texture mapping issue, making it difficult to meet the requirements for intuitive representation and traceable archiving of surface defects during the operation and maintenance phase. Furthermore, existing image-based texture acquisition methods largely rely on dense reconstruction (MVS) to first generate a complete mesh and appearance, followed by multi-view texture fusion. This process heavily depends on high-resolution, multi-view depth, and global optimization, resulting in long computation time, high memory and storage consumption, sensitivity to highly reflective metals and repetitive textures, and susceptibility to occlusion at node intersections, leading to holes and noise. Moreover, multi-source splicing easily introduces stitching marks and color inconsistencies. More importantly, these general processes lack adaptation and constraints for the geometric features of "parameterized circular tubes + intersecting Boolean geometry," making it difficult to balance efficiency and stability on an engineering scale. Summary of the Invention
[0009] The purpose of this invention is to address the shortcomings of the prior art by providing a texture reconstruction method for intersecting welded joints of circular pipes in steel structures. This method completes mesh preparation and texture assignment based on sparse / semi-dense information, balances geometric fitting robustness and texture continuity in complex intersecting welded joint scenarios, reduces multi-source fusion overhead, and is suitable for applications such as engineering inspection, BIM / digital twins, and defect record retention.
[0010] The objective of this invention is achieved through the following technical solutions: A texture reconstruction method for intersecting welded joints of circular pipes in steel structures, characterized by the following steps: S1. Using initial images taken from multiple angles of the intersecting welded joint of the steel structure circular pipe, reconstruct the sparse point cloud of the steel structure circular pipe and calculate the camera pose corresponding to each angle. S2. Perform parametric geometric modeling of the steel structure circular tube based on the sparse point cloud to obtain a cylindrical model of the steel structure circular tube based on the sparse point cloud. S3. Process the cylindrical model of the steel structure tube to generate a semi-dense point cloud on the surface of the steel structure tube at the intersecting weld joint. S4. Perform visualization filtering on the semi-dense point cloud; S5. Perform texture mapping based on the semi-dense surface point cloud after visualization filtering.
[0011] In step S1, a sequence of images covering the surface of the steel structure circular tube is acquired; the sequence of images is processed based on a semantic segmentation model to generate a foreground region mask for the steel structure circular tube; based on the foreground region mask, image feature matching and incremental motion recovery structure calculation are performed, and sparse 3D point cloud and camera pose are solved simultaneously.
[0012] In step S2, adaptive clustering analysis is performed on the sparse point cloud to obtain the total number of sparse point clouds and the three-dimensional boundary size; initial geometric estimation of the cylinder is performed based on the total number of sparse point clouds and the three-dimensional boundary size; the cylinder is fitted and energy evaluated based on least squares optimization; and the cylinder is extracted based on the fitting result.
[0013] By combining global random sampling and cluster sampling, a hybrid sampling strategy is adopted, which samples within a randomly selected cluster with a 70% probability and samples the global point cloud with a 30% probability, to obtain the sampling point set used to fit the initial value of the cylinder.
[0014] When extracting a single cylinder, iterative optimization is performed with the goal of maximizing energy. The calculation terminates when the proportion of interior points reaches a certain target threshold or when there is no energy improvement after several consecutive iterations. When extracting multiple cylinders, after determining a set of optimal cylinder parameters each time, the interior points corresponding to the cylinder are removed from the set of remaining points. Adaptive clustering analysis, initial geometric estimation of the cylinder, and cylinder fitting and energy evaluation based on least squares optimization are repeatedly performed on the remaining points until the maximum number of cylinders is extracted or the proportion of remaining points is lower than a preset threshold.
[0015] In step S3, based on the cylindrical model, the effective length extension and sampling parameters of the cylinder based on the sparse point cloud are designed; semi-dense surface sampling and normal vectors are generated; Boolean difference sets of the intersecting regions of the steel structure circular tubes are screened out to obtain a semi-dense point cloud of the steel structure circular tube surface containing the normal vector.
[0016] For all sampling points of the cylinder, perform vectorized inside / outside determination for all other cylinders, calculate the distance of all sampling points to the axis of other cylinders, and when the distance of a point to the axis of a certain cylinder is less than a preset threshold, determine that the point belongs to the inside of other cylinders and remove it, only retaining the points that are not inside any other cylinders.
[0017] In step S4, surface mesh reconstruction is performed based on the semi-dense point cloud of the steel structure circular tube surface containing normal vectors. For each reconstructed view, a multi-view depth map with adaptive resolution is generated, and only the depth map of the foreground region is retained. Visibility judgment based on depth consistency is performed on the depth map of the foreground region. Visual filtering and color assignment are performed on the depth map of the foreground region after visibility judgment to obtain the semi-dense outer surface point cloud.
[0018] For any point in a semi-dense point cloud, count the number of visible viewpoints for each point; set a minimum threshold for the number of visible viewpoints, retain only points with a pre-defined number of visible viewpoints greater than the minimum threshold and form a set of visible views, calculate the pixel coordinates based on the camera's intrinsic and extrinsic parameters and read the pixel colors from the corresponding original images, and take the average value of the colors read from each view to obtain the color of that point.
[0019] In step S5, the semi-dense outer surface point cloud obtained in step S4 is reconstructed into a mesh and divided into several triangular patches; for each triangular patch, a single imaging view is assigned as the texture source and texel writing is completed.
[0020] The advantages of this invention are: 1) Lightweight and efficient: Bypassing full-scene dense reconstruction (MVS) and complex multi-source fusion, it completes mesh preparation and texture assignment based on sparse / semi-dense information, significantly reducing computation and storage overhead, and is suitable for batch node processing. 2) High project usability: The process is simple and can be integrated with BIM / digital twin and inspection systems, which facilitates rapid on-site modeling and record keeping.
[0021] 3) It has advantages such as low computational overhead, robustness to high-contrast and occluded scenes, easy engineering integration and batch processing.
[0022] 4) It is applicable to the three-dimensional visualization modeling, defect visualization archiving and operation and maintenance display of various hollow circular cross-section (CHS) components such as space grid structures, large steel structure roofs, bridge steel trusses, and curtain wall support systems, and has a wide range of applications. Attached Figure Description
[0023] Figure 1 This is a flowchart of the present invention; Figure 2 These are initial images taken from multiple angles of the intersecting welded joint of the steel structure circular tube in this invention; Figure 3 This is a foreground region segmentation diagram in this invention; Figure 4 This is a sparse point cloud and camera pose diagram in this invention; Figure 5 This is a diagram showing the effect of fitting the cylindrical surface of three cylinders and their corresponding final interior points in this invention; Figure 6 This is a semi-dense point cloud and normal vector map of the circular tube surface at the intersecting weld node generated in this invention. Figure 7 These are the depth maps corresponding to each original image in this invention; Figure 8 This is a visualization result of the semi-dense point cloud, including filtering and color assignment. Figure 9 This is a diagram showing the final mesh division effect in this invention; Figure 10 This is the final texture reconstruction result in this invention. Detailed Implementation
[0024] The features and other related features of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments, so as to facilitate understanding by those skilled in the art: Example: Figure 1 As shown, the texture reconstruction method for the intersecting weld joint of steel structure circular pipes in this embodiment includes the following steps: Step 1: Sparse point cloud reconstruction: (1) Use a camera to acquire initial images of the intersecting welded joint of the steel structure circular pipe from multiple angles, ensuring that the images cover the main area of the component surface, such as... Figure 2 As shown.
[0025] (2) The pre-trained SAM (Segment Anything Model) model is used to extract the surface region (foreground region) of the circular tube in each image, such as... Figure 3 As shown.
[0026] (3) Within the foreground region mask, local feature extraction is performed on the image. The SIFT algorithm is preferred for extracting image feature points. Feature point pairing between images is completed based on descriptor distance matching. An incremental reconstruction method is used to recover the sparse point cloud of the steel structure circular pipe member and calculate the global camera pose. A total of 4023 sparse points and 9 camera poses are included, such as... Figure 4 As shown.
[0027] Step 2: Parametric geometric modeling: (1) Adaptive clustering analysis: Obtain the total number of point clouds and the 3D boundary dimensions; when the total number of point clouds > 1000, use K-Nearest Neighbors (KNN) with k=10 to estimate the average distance (ds) between points (average of the 10 nearest neighbors of each point and then global average), and adaptively set the DBSCAN parameters accordingly: neighborhood radius (ds=max(3ds,0.05), minimum number of samples min_samples=max(10, ), where To round down; when the total number of point clouds is less than 1000, set ds=0.1 and min_samples=5. Execute DBSCAN with the above parameters to obtain cluster labels. ,in =-1 is marked as noise and removed, only valid clusters are retained to form a subset. If the DBSCAN output has insufficient clusters or a large proportion is judged as noise, it degenerates into K-means and adaptively sets the cluster number K=min(10, ); When the total number of point clouds < K, take K = max(1, ). Each of the finally output subsets serves as a candidate grouping for subsequent sampling and cylinder fitting to improve sampling representativeness and robustness.
[0028] (2) Cylindrical geometric initial value estimation: Combining two methods of global random sampling and in-cluster sampling, a hybrid sampling strategy is adopted. With a 70% probability, sample within a randomly selected cluster, and with a 30% probability, sample from the global point cloud to obtain a set of sampling points for fitting the initial value of the cylinder. The number of samples per time is 10; in the sampling point set, based on principal component analysis (PCA), solve the principal direction vector of the point cloud as the initial value of the cylinder axis direction ; Calculate the centroid of the sampling points as the initial passing point of the axis ; Estimate the initial radius value based on the variance of the point cloud along the axis direction .
[0029] (3) Cylinder fitting based on least squares optimization and energy evaluation: Use the nonlinear least squares method to optimize the residual function of the sampling points in (2). For any point and the cylinder parameters , the residual is defined as the difference between the perpendicular distance from the point to the cylinder axis and the radius: , and minimize the sum of squared residuals of the sampling point set to obtain the local optimal parameters. After solving the temporary cylinder parameters for each sampling, calculate the distance from the points to the cylinder surface on all the remaining points, determine the inliers with a distance threshold of 0.01, and obtain the number of inliers and the average residual distance of the inliers , and calculate the energy function of the fitted cylinder according to Equation ; (4) Sequential extraction of multiple cylinders For the extraction of a single cylinder, repeat the process from (1) to (3), and iteratively optimize with the goal of maximizing the energy E. When the inlier ratio reaches the target threshold of 0.7 or there is no energy improvement after several consecutive (50) iterations, terminate the calculation in advance. Use the final parameters to recalculate the distance from all points to the cylinder surface, and determine the final inlier set with a review threshold of 0.03; when the point cloud contains multiple intersecting circular tubes, after determining a set of optimal cylinder parameters each time, remove the inliers corresponding to the cylinder from the remaining point set, and repeat steps (1) to (3) on the remaining points until the maximum number of cylinders is extracted or the remaining point ratio is lower than the preset threshold of 2%; if the final fitted inlier ratio of a single cylinder is lower than the lowest threshold of 5%, also terminate further extraction.
[0030] Figure 5The image shows the effect of fitting three cylinders to the cylindrical surface and their respective final interior points.
[0031] Step 3: Generation of semi-dense point clouds: (1) Effective length extension of cylinder and setting of sampling parameters based on sparse point cloud: For each cylinder obtained in the parametric geometric modeling stage, the final set of interior points will be obtained in step (3-4). The projection range of the cylinder on the axis of the cylinder is calculated. The effective length is obtained by extending the cylinder's length along its boundary, and the axial center of the cylinder is also determined. The maximum radius of the three cylinders is 1.404, and the longest axis is 3.384. The sample counts corresponding to the maximum radius and the longest axis are set to 360 and 250 respectively. For the remaining cylinders, the corresponding sample counts are calculated proportionally according to the ratios of the radius, axis length, maximum radius, and longest axis. and .
[0032] (2) Semi-dense surface sampling and normal vector generation For the For each cylinder, construct a structure with its axis vector. Orthogonal two-dimensional orthogonal basis Take evenly along its axis Axial sample points ,exist Take evenly from the top Corner sample points According to the formula Generate surface points according to the formula Generate the unit normal vector for each point.
[0033] (3) Boolean difference set screening of intersecting regions: For the All sampling points generated by the cylinder For all other cylinders, a vectorized inside / outside determination is performed, specifically by calculating the distance from all sampling points to the axes of other cylinders. When the distance from a point to the k-th cylinder axis is less than When this point is found to be an interior point of another cylinder, it is determined and discarded. To determine the tolerance, take Only retain the points that are not inside any other cylinder, and take the union of all cylinders to obtain the semi-dense outer surface point cloud of the intersecting weld node.
[0034] Figure 6 The image shows the semi-dense point cloud and normal vector on the surface of the circular tube at the intersecting weld joint. The few dark points in the image are sparse reconstructed point clouds.
[0035] Step 4: Visualizing and filtering semi-dense point clouds: (1) Surface mesh reconstruction: Based on the reconstruction of a triangular mesh from a semi-dense point cloud containing normal vectors, Poisson reconstruction was used with the reconstruction depth and scale parameters set (scale factor s=0.125), resulting in a total of 288,667 triangular patches.
[0036] (2) Resolution-adaptive multi-view depth map generation: For each reconstructed view, the camera model and parameters (including the intrinsic parameter matrix K formed by the focal length and principal point, and the pose T determined by quaternions and translation) of that view are read from the sparse reconstruction results; the image resolution is scaled synchronously with K according to the scaling factor s=0.125, and the depth map D is initialized as follows: ;Will The middle triangular face is projected onto the image plane, and the depth of the three vertices is interpolated using the centroid coordinates and written to the nearest depth using the Z-buffer strategy to obtain the scaled depth map of the view. The scaled depth map is upsampled to the original resolution and filtered by combining it with the foreground region mask map from the first step, retaining only the foreground region depth map.
[0037] Figure 7 The depth maps corresponding to each original image are displayed.
[0038] (3) Visibility judgment based on depth consistency: For any point x in a semi-dense point cloud, the pixel coordinates (a, b) and camera coordinate depth are obtained by projecting the data onto the cloud using the camera's intrinsic and extrinsic parameters in each view. When (a,b) is located on the image plane and the corresponding foreground mask is valid, and simultaneously satisfies When a point is visible in a view, record that the point is visible and count the number of visible viewpoints for each point. .
[0039] (3) Visual filtering and color assignment: Set the minimum visible view threshold to 1, and retain only those that satisfy... Points constitute ;right For each point in the set of visible views, the pixel coordinates are calculated based on the camera's intrinsic and extrinsic parameters, and the pixel color is read from the corresponding original image. The average value of the colors read from each view is then taken to obtain the color of that point.
[0040] Figure 8 The results of visualization filtering and color assignment for semi-dense point clouds are shown.
[0041] Step 5, Texture Mapping: (1) Grid reconstruction: Based on the semi-dense outer surface point cloud obtained in step four, repeat the first step of step four to obtain the final mesh.
[0042] Figure 9 The final grid division is shown.
[0043] (2) Texture mapping: For each triangular facet of the target mesh, only a single imaging view is assigned as the texture source and texel writing is completed, specifically including: View assignment: Based on the point-level visibility analysis in step four, select the set of views that simultaneously observe all vertices of the face; for each view Calculate the comprehensive evaluation function , Pick .in, The unit normal vector of the surface. For the center of the dough, Center the camera in view I; select The view with the highest score is used as the sole texture source for that facet.
[0044] Texture mapping and seam mitigation: The projected pixel coordinates of the three vertices of the face in the view are normalized to UV coordinates according to the image width and height, and texel sampling and writing are performed within the UV triangle using perspective-consistent centroid interpolation; to mitigate the visual seams caused by single-source textures, within the face's own UV domain, the distance from the boundary is not more than... Separable Gaussian feathering is applied to the narrowband, r=2 pixels, and the standard deviation of the Gaussian kernel is set. =1.
[0045] Figure 10 The final texture reconstruction result is shown.
[0046] Although the above embodiments have described the concept and embodiments of the present invention in detail with reference to the accompanying drawings, those skilled in the art will recognize that various improvements and modifications can still be made to the present invention without departing from the scope of the claims, and therefore will not be elaborated here.
Claims
1. A texture reconstruction method for a steel tubular chord intersecting welded joint, characterized in that: The texture reconstruction method includes the following steps: S1. Using initial images taken from multiple angles of the intersecting welded joint of the steel structure circular pipe, reconstruct the sparse point cloud of the steel structure circular pipe and calculate the camera pose corresponding to each angle. S2. Perform parametric geometric modeling of the steel structure circular tube based on the sparse point cloud to obtain a cylindrical model of the steel structure circular tube based on the sparse point cloud. S3. Process the cylindrical model of the steel structure tube to generate a semi-dense point cloud on the surface of the steel structure tube at the intersecting weld joint. S4. Perform visualization filtering on the semi-dense point cloud; S5. Perform texture mapping based on the semi-dense surface point cloud after visualization filtering.
2. The method according to claim 1, wherein: In step S1, a sequence of images covering the surface of the steel structure circular tube is acquired; the sequence of images is processed based on a semantic segmentation model to generate a foreground region mask for the steel structure circular tube; based on the foreground region mask, image feature matching and incremental motion recovery structure calculation are performed, and sparse 3D point cloud and camera pose are solved simultaneously.
3. The method according to claim 1, wherein: In step S2, adaptive clustering analysis is performed on the sparse point cloud to obtain the total number of sparse point clouds and the three-dimensional boundary size; initial geometric estimation of the cylinder is performed based on the total number of sparse point clouds and the three-dimensional boundary size; the cylinder is fitted and energy evaluated based on least squares optimization; and the cylinder is extracted based on the fitting result.
4. The method according to claim 3, wherein: By combining global random sampling and cluster sampling, a hybrid sampling strategy is adopted, which samples within a randomly selected cluster with a 70% probability and samples the global point cloud with a 30% probability, to obtain the sampling point set used to fit the initial value of the cylinder.
5. The texture reconstruction method for intersecting welded joints of circular pipes in steel structures according to claim 3, characterized in that: When extracting a single cylinder, iterative optimization is performed with the goal of maximizing energy. The calculation terminates when the proportion of interior points reaches a certain target threshold or when there is no energy improvement after several consecutive iterations. When extracting multiple cylinders, after determining a set of optimal cylinder parameters each time, the interior points corresponding to the cylinder are removed from the set of remaining points. Adaptive clustering analysis, initial geometric estimation of the cylinder, and cylinder fitting and energy evaluation based on least squares optimization are repeatedly performed on the remaining points until the maximum number of cylinders is extracted or the proportion of remaining points is lower than a preset threshold.
6. The texture reconstruction method for intersecting welded joints of circular pipes in steel structures according to claim 1, characterized in that: In step S3, based on the cylindrical model, the effective length extension and sampling parameters of the cylinder based on the sparse point cloud are designed; semi-dense surface sampling and normal vectors are generated; Boolean difference sets of the intersecting regions of the steel structure circular tubes are screened out to obtain a semi-dense point cloud of the steel structure circular tube surface containing the normal vector.
7. The texture reconstruction method for intersecting welded joints of circular pipes in steel structures according to claim 6, characterized in that: For all sampling points of the cylinder, perform vectorized inside / outside determination for all other cylinders, calculate the distance of all sampling points to the axis of other cylinders, and when the distance of a point to the axis of a certain cylinder is less than a preset threshold, determine that the point belongs to the inside of other cylinders and remove it, only retaining the points that are not inside any other cylinders.
8. The texture reconstruction method for intersecting welded joints of round steel bars in steel structures according to claim 1, characterized in that: In step S4, surface mesh reconstruction is performed based on the semi-dense point cloud of the steel structure circular tube surface containing normal vectors. For each reconstructed view, a multi-view depth map with adaptive resolution is generated, and only the depth map of the foreground region is retained. Visibility judgment based on depth consistency is performed on the depth map of the foreground region. Visual filtering and color assignment are performed on the depth map of the foreground region after visibility judgment to obtain the semi-dense outer surface point cloud.
9. A texture reconstruction method for intersecting welded joints of round steel bars in steel structures according to claim 8, characterized in that: For any point in a semi-dense point cloud, count the number of visible viewpoints for each point; set a minimum threshold for the number of visible viewpoints, retain only points with a pre-defined number of visible viewpoints greater than the minimum threshold and form a set of visible views, calculate the pixel coordinates based on the camera's intrinsic and extrinsic parameters and read the pixel colors from the corresponding original images, and take the average value of the colors read from each view to obtain the color of that point.
10. The texture reconstruction method for intersecting welded joints of round steel bars in steel structures according to claim 1, characterized in that: In step S5, the semi-dense outer surface point cloud obtained in step S4 is reconstructed into a mesh and divided into several triangular patches; for each triangular patch, a single imaging view is assigned as the texture source and texel writing is completed.