Method for Determining Geometric Features of Bridge Components Based on Hybrid Model

Abstract

The present invention provides a method for determining the geometric features of bridge components based on a hybrid model. First, a bridge component recognition model is constructed and trained, and then this model is used to recognize and segment the point cloud data of bridge components. Next, multi-scale features are extracted to segment the geometric boundaries and skeletons, and they are fused with the wireframe model to construct a dual-mode fusion representation model containing point clouds and wireframes. Finally, based on the above model and the point cloud data, the distance from each point to the nearest edge is calculated to generate a distance map, and the distance map is used for precise measurement of the dimensions of bridge components to ensure the accuracy of butt joint of each segment. The present invention uses region growing segmentation and a "point cloud + wireframe" hybrid model to simplify data processing and improve computational efficiency. By using Laplacian eigenmap and Alpha shape algorithm, high-precision dimension measurement is achieved, which is applicable to the detection of bridge components under complex geometric conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge construction, and particularly relates to a method for determining geometric features of bridge components based on a hybrid model. Background Art

[0002] Segment precast components are an important part of the industrialized construction of bridges. The actual quality of the components after being produced in the factory has a crucial impact on the on-site construction. Segment precast components usually adopt the construction method of on-site assembly, which has higher requirements for the alignment control and splicing error of the assembled segments. Any assembly error of a segment will lead to problems in the quality and efficiency of the final completed bridge. To ensure that the quality of segment precast components meets the standards, reduce the construction assembly error and construction risks, spatially related precast components usually need to be temporarily assembled before leaving the factory, that is, pre-assembled, so as to provide guiding opinions on the processing and linear control of the components before actual assembly. Pre-assembly requires additional different temporary lifting equipment and safety protection measures. Although this method is closer to the actual situation, it has the disadvantages of high safety hazards, long pre-assembly construction period, large site occupation, high costs of jigs and labor, etc. In addition, for complex environments such as mountainous areas and cross-(river)sea areas, due to site and equipment conditions restrictions, the pre-assembly of (extra) large-span bridge components cannot be fully realized, and the unit modular method is often used for pre-assembly, which increases the project cost, delays the construction progress, and to a certain extent affects the quality and efficiency of actual assembly.

[0003] With the demand for the application of segment precast components in the industrialized construction of bridges and the continuous development of science and technology, the virtual assembly theory and method with higher assembly efficiency and quality of bridge components gradually shows application prospects in bridge construction.

[0004] However, there are still the following deficiencies in the technical reserve research for realizing the automatic recognition of bridge precast components and then obtaining their finite element models:

[0005] Traditional methods have some limitations in the recognition and segmentation of bridge components. Because they rely on manually designed features, they cannot comprehensively capture complex shapes and details, resulting in low recognition accuracy. The insufficient utilization of spatial information also leads to the inability to effectively process the relative position relationship between components; and they perform poorly in the presence of noise, occlusion or data loss, with low robustness; they are prone to ignoring the context relationship between components and are prone to misrecognition or omission.

[0006] In terms of point cloud data processing, in the process of simulating the assembly matching of adjacent segment point cloud data, the selection of the alignment area and a large number of iterative calculations are involved, which is prone to the unnecessary bigness of data parameters and cannot accurately represent the outer contour of bridge components while performing fine-grained modeling on the key parts during the docking assembly process.

[0007] In the inspection of the dimensions of bridge precast components, existing traditional manual methods face great difficulties in large precast components, especially in the inspection of the linearity of prestressed ducts. Since prestressed ducts are usually buried inside or on the surface of components and have complex linearity, manual inspection has low accuracy and is prone to errors. In addition, manual inspection requires a large amount of time and energy, with high labor intensity. Especially in large components, the inspection process takes a long time and is inefficient. As the scale and complexity of components increase, the limitations of traditional methods become more obvious, and there is an urgent need for more efficient and intelligent technical means to improve inspection accuracy and efficiency. Summary of the Invention

[0008] The purpose of the embodiments of the present invention is to provide a method for determining the geometric features of bridge components based on a hybrid model, which solves the problems of large safety hazards, high costs, low efficiency, and limited site in the pre-assembly process of traditional bridge precast components, as well as the technical problems of insufficient component recognition accuracy, low point cloud data processing efficiency, and difficult manual inspection of key parts such as prestressed ducts in virtual assembly technology.

[0009] To solve the above technical problems, the technical solution adopted by the present invention is a method for determining the geometric features of bridge components based on a hybrid model, which is specifically carried out according to the following steps:

[0010] S1. Construct and train a bridge component recognition model, and then use this model to recognize and segment the point cloud data of bridge components;

[0011] S2. Extract and segment the geometric boundaries and skeletons through multi-scale features, and fuse them with the wireframe model to construct a dual-mode fusion representation model containing point clouds and wireframes;

[0012] S3. Determine the dimensions of bridge components.

[0013] Further, the bridge component recognition model in S1 includes a relative position encoding module and a spatial attention module;

[0014] The relative position encoding module includes an input module, a feature extraction module, a global feature extraction module, a normalization module, an embedding module, a feature preprocessing module, and an output module connected in sequence;

[0015] The input module includes a point feature matrix of N×d, where N is the number of points in the point cloud of the bridge component, and d is the feature dimension of each point;

[0016] The feature extraction module includes three SA layers, and each SA layer includes a PointNet module. The PointNet module includes an MLP and a global feature extraction module; the MLP is used for non-linear transformation of the features of each point, and the global feature extraction module is used for maximum pooling of the point features in the local area to extract global features;

[0017] A normalization module that normalizes features;

[0018] An embedding module that uses the normalized features as embedding representations and encodes relative position information as a vector of a fixed dimension;

[0019] A feature preprocessing module that preprocesses the embedded features through two layers of MLP;

[0020] An output module that includes a final feature matrix of N×d;

[0021] The spatial attention module includes a self-attention mechanism for implementing the calculation process of attention weights.

[0022] Furthermore, the self-attention mechanism is specifically:

[0023] The input feature map is , where is the height, is the width, is the number of channels, represents the set of real numbers;

[0024] First, by performing a linear transformation on the input feature map, query vector , key vector and value vector are calculated:

[0025] ;

[0026] Among them, is the learned weight matrix; is the feature dimension of each point;

[0027] Calculate the attention weights. First, calculate the similarity score and : :

[0028] ;

[0029] Among them, is the result of scaled dot product; The function is used to normalize the scores; T is the transpose symbol;

[0030] Finally, apply the attention weights to the value vector to obtain the final output :

[0031] .

[0032] Further, the training process of the bridge component recognition model is specifically as follows:

[0033] (1) Process the bridge component point cloud data in batches. Each batch contains multiple points, and each point contains a coordinate position (x, y, z) and additional feature values for each point; the additional feature values include normal vectors, surface directions, curvatures, and concavities and convexities.

[0034] (2) Encode the batched bridge component point cloud data.

[0035] (3) Standardize the encoded bridge component point cloud data.

[0036] (4) Augment the bridge component point cloud data, including rotation, flipping, and scaling.

[0037] (5) Insert a spatial attention module between the first and second SA layers of the bridge component recognition model; input the augmented point cloud data into the bridge component recognition model to extract local features and generate a global feature vector representing the entire point cloud.

[0038] Each execution of the SA layer includes three steps:

[0039] a. Sampling: Use farthest point sampling to select a set of center points from the point cloud.

[0040] b. Grouping: Based on each center point, select its K nearest neighbor points to construct a local region. Each local region contains a center point and its neighbor points.

[0041] c. Feature extraction: Use a PointNet module to extract features from each local region.

[0042] (6) The global feature vector is further processed through an MLP layer to output the class probability distribution of the object.

[0043] (7) Use the loss function of cross-entropy loss to train the model by minimizing the difference between the predicted class and the true class.

[0044] Further, the segmentation process of the bridge component is specifically as follows: Randomly select multiple seed points in the bridge component point cloud data after being recognized by the bridge component recognition model as the starting points of segmentation; Starting from the initial seed points, gradually check whether the adjacent unlabeled points meet the normal consistency similarity criterion; If the criterion is met, add the point to the current region and continue to expand; Stop until there are no eligible points to add, and obtain the preliminary segmentation region.

[0045] After the expansion of an area is completed, a new seed point is selected to continue the segmentation until the preset segmentation goal is reached; finally, the point cloud is divided into several non-overlapping regions or categories, and the category to which each point belongs is output to complete the segmentation of bridge components.

[0046] Furthermore, the construction process of the dual-mode fusion representation model of the point cloud and the wireframe is specifically as follows:

[0047] S201. Extract the geometric boundary information of the segmented component parts;

[0048] S202. Generate a skeleton based on the extracted boundary information, and the skeleton includes the geometric center and skeleton lines of the component;

[0049] S203. Combine the generated skeleton information with the bridge component point cloud data to perform fine-grained modeling on the key parts during the assembly process;

[0050] S204. Smooth and simplify the fine-grained modeling results; obtain a dual-mode fusion representation model including point cloud and wireframe.

[0051] Furthermore, the specific process of the fine-grained modeling in S203 is as follows: Based on the generated skeleton information combined with the bridge component point cloud data, mark the key point coordinates at the branch points and end points of the component, and extract their normal vectors and curvatures; then, register and align the skeleton information with the point cloud data to ensure that the two are in the same coordinate system, and then extract the local point cloud data of the key points; preprocess the local point cloud data to remove noise and outliers; perform surface fitting on the local point cloud to generate a fine-grained geometric model, and generate a continuous surface based on the point cloud normal vector; add detailed information at the key parts, extract boundary points from the point cloud to strengthen the geometric representation at the seams; check whether the fine-grained model is consistent with the original point cloud, evaluate the accuracy of the fine-grained geometric model using the distance error from the point to the model, and if the accuracy is insufficient, re-extract the local point cloud and repeat the above steps; finally, output the specific parameters of the 3D model.

[0052] Furthermore, the specific process of S3 is as follows:

[0053] S301. Map the 3D point cloud data of the dual-mode fusion representation model including point cloud and wireframe to 2D; specifically, first construct a graph to represent the local relationship in the 3D point cloud data of the dual-mode fusion representation model including point cloud and wireframe, and then obtain the low-dimensional representation by calculating the eigenvectors of the Laplacian matrix of the graph;

[0054] S302. Extract the edge information of the bridge component;

[0055] S303. Determine the size of the bridge component.

[0056] Further, the specific process of S302 is as follows: Extract the edge lines of the component based on the Alpha shape algorithm, and flexibly adjust the extraction accuracy by setting different Alpha values; Smooth the extracted edge data, and use the adjacent point connection algorithm to repair the missing edge connections to ensure the integrity and continuity of the edge model, and finally obtain the optimized edge model.

[0057] Further, the specific process of S303 is as follows: Use the KD-Tree data structure to accelerate the calculation of the distance from each point in the point cloud to the nearest edge, generate a distance map, and accurately measure the component size based on the distance map.

[0058] Compared with the prior art, the beneficial effects of the present invention include the following points:

[0059] 1. The present invention is a fast and reliable geometric feature detection algorithm, which has the advantages of high precision, high recognition efficiency, accurate segmentation type, short time consumption, and accurate key dimensions, compared with the low efficiency and long calculation time of general detection algorithms.

[0060] 2. The efficient recognition and segmentation method used in the present invention has accuracy and efficiency. The improved bridge component recognition model established based on PointNet++ combined with the spatial attention module and position encoding can capture complex shapes and details, and the model obtained through effective training sets and test sets has high credibility. The segmentation model algorithm based on region growing can effectively extract the feature information of components, make full use of spatial information to effectively process the relative positions between components, and achieve the purpose of separating surface regions based on algorithm learning.

[0061] 3. The "point cloud + wireframe" hybrid model proposed in the present invention has high applicability. Compared with the traditional point cloud model, the data is unnecessarily large, resulting in slow operation and low efficiency. When completing the necessary link of converting point cloud data into a parametric model, the data processing and iterative calculation amount are simplified. Extract the geometric boundaries and skeleton information of the segmented component parts through the wireframe model, and fuse the high-precision geometric information of the point cloud with the wireframe model.

[0062] 4. The high-precision detection method for the dimensions of precast bridge components proposed in the present invention is based on the algorithm established by the "point cloud + wireframe" model. The Laplace eigenmap method is used to simplify complex three-dimensional data into two-dimensional images that are easy to process, accurately extract the edge features of components in combination with the Alpha shape algorithm, and then calculate the distance from each pixel to the nearest edge through the distance transformation technology to achieve high-precision measurement of component dimensions. It can overcome the detection problems of bridge components under complex geometric conditions. Description of the Drawings

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0064] Figure 1 is the technical roadmap of this embodiment;

[0065] Figure 2 is the schematic diagram of the relative position encoding module;

[0066] Figure 3 is the recognition effect diagram of the recognition model of the bridge components in this embodiment;

[0067] Figure 4 is the recognition process of the recognition model of the bridge components in this embodiment;

[0068] Figure 5 is the segmentation result based on region growing in this embodiment. Specific Embodiments

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0070] Existing segment precast components usually adopt the construction method of on-site assembly, which has higher requirements for the linear control of the assembled segments and the splicing error. Any error in segment assembly will lead to problems in the quality and efficiency of the final completed bridge.

[0071] This embodiment proposes a method for determining the geometric features of bridge components based on a hybrid model. First, PointNet++ is used as the basic framework to extract the multi-scale geometric features of bridge components, and the extracted features are processed using the region-growing-based segmentation model algorithm to achieve precise segmentation of each key part of the components. On the basis of the point cloud data of bridge components, when completing the necessary steps of converting the point cloud data into a parametric model, by simplifying data processing and reducing the amount of iterative calculations, the high-precision geometric information of the point cloud is fused with the wireframe model to form a 'point cloud + wireframe' hybrid model containing rich geometric information. Finally, a geometric feature detection algorithm for bridge components of the 'point cloud + wireframe' hybrid model is proposed to achieve high-precision measurement of component dimensions. It can overcome the limitations of traditional measurement techniques under complex geometric conditions and provide a new technical path for the precise detection of bridge precast components.

[0072] In some specific embodiments, such as Figure 1 , the method for determining the geometric features of bridge components based on the hybrid model mainly includes the efficient recognition and segmentation of bridge components, the construction of a "point cloud + wireframe" hybrid model, and the high-precision detection of the dimensions of bridge components.

[0073] S1. Efficient recognition and segmentation of bridge components

[0074] S101. Construct and train a bridge component recognition model, and use the recognition model to recognize the collected point cloud data;

[0075] PointNet++ is a deep learning architecture for processing point cloud data, mainly applied to tasks such as 3D object recognition, classification, and segmentation. Different from traditional convolutional neural networks (CNNs), PointNet directly processes the original point cloud data without meshing or rasterization, avoiding the risk of point cloud data loss. Different from the original PointNet, PointNet++ enhances the understanding of local details by introducing a hierarchical structure, enabling it to capture local features of point clouds more effectively. This embodiment is based on the traditional PointNet method, combined with a spatial attention mechanism (SAM) module and a relative position encoding module, to improve the accuracy of bridge component feature extraction and segmentation.

[0076] The bridge component recognition model based on PointNet++ in this embodiment includes a relative position encoding module and a spatial attention module;

[0077] In some specific embodiments, the relative position encoding module processes the point cloud data based on the relative coordinates of the point cloud, matches the dimensions, and embeds or adds them as initial features into the neural network. Relative position encoding is achieved by encoding the relative position coordinates between a point and its neighbor points:

[0078] ;

[0079] Among them, represents the position encoding vector of point , , respectively represent the central point coordinates of point and its neighbor point , represents the encoding function, which is used to convert into an encoded form to help the model aggregate local features and understand context information.

[0080] Before encoding, the K-nearest neighbor algorithm is used to find the near point The points are divided into their neighboring points. In this embodiment, the encoding process can be regarded as constructing a PointNet module to extract relative position information through three-layer MLP and max pooling. Then, point convolution is used to match the dimensions. Subsequently, the encoding is normalized and used as an initial feature embedding or added to the neural network, and then two-layer MLP is used for feature preprocessing for aggregation.

[0081] In some specific embodiments, the structure of the relative position encoding module is as Figure 2 shown, including an input module, a feature extraction module, a global feature extraction module, a normalization module, an embedding module, a feature preprocessing module, and an output module connected in sequence.

[0082] In some possible embodiments, the input module includes a point feature matrix of N×d, where N is the number of points in the bridge component point cloud and d is the feature dimension of each point (coordinates, normal vector); the feature extraction module performs a non-linear transformation on the input features through three-layer MLP (multi-layer perceptron) to extract high-level features of the points; the global feature extraction module performs max pooling on the features to obtain global features (relative position information); the normalization module normalizes the features to improve the training stability and convergence speed of the model; the embedding module uses the normalized features as an embedding representation to encode the relative position information into a vector of a fixed dimension; the feature preprocessing module preprocesses the embedded features through two-layer MLP; the output module includes a final feature matrix of N×d for subsequent output tasks.

[0083] In some specific embodiments, the spatial attention mechanism module includes a self-attention mechanism for implementing the calculation process of attention weights; the self-attention mechanism can adaptively learn the dependency relationships between each position according to the input feature map, so as to achieve importance weighting of different positions.

[0084] In some specific embodiments, the self-attention mechanism is specifically as follows:

[0085] The input feature map is , where is the height, is the width, and is the number of channels. First, by performing a linear transformation on the input feature map, calculate and , and the formulas are shown in Equations (1)-(3).

[0086] (1)

[0087] (2)

[0088] (3)

[0089] Where: is the learned weight matrix; is the feature dimension of each point.

[0090] To calculate the attention weights, first calculate the and similarity score between them, using dot-product attention;

[0091] ;

[0092] Where: is the result of scaled dot product; The function is used to normalize the score.

[0093] Finally, apply the attention weights to , and the output formula of the weighted sum is shown in (4).

[0094] (4)

[0095] Through the above formula, the spatial attention module can calculate the attention weights of each position according to the input feature map, so as to realize the weighted fusion of feature representations at different positions and improve the model's ability to focus on local features.

[0096] (1) Process the collected bridge component point cloud data batch by batch. Each batch contains multiple points, and each point contains coordinate positions (x, y, z) and its additional feature values (including normal vector, surface direction, curvature, concavity and convexity, etc.).

[0097] (2) Encode the data according to the relative position encoding module before inputting the data.

[0098] (3) Normalize the processed point cloud in (2). First, translate the point cloud to the origin, and then scale it to a unified size range (such as a unit cube).

[0099] (4) Data augmentation processing. Before training, process the point cloud data through measures such as rotation, flipping, and scaling to improve the robustness of the model.

[0100] (5) The input point cloud data will pass through a multi-layer perceptron (MLP) to extract local features from each point, and then process the point cloud step by step through a hierarchical structure. At each layer, first perform clustering or neighborhood selection of the local point set (such as k-nearest neighbor or sphere sampling), and then extract local features through a local network. Then aggregate the point features through max pooling to generate a global feature vector representing the entire point cloud.

[0101] In some specific embodiments, the model of the present application based on PointNet++ further includes three SA layers, which gradually downsample and extract features; before the point cloud data enters the first SA layer and before it enters the second SA, it enters an attention weight processing module, which can adaptively adjust the feature weights according to the importance of different positions in the data, so as to process the point cloud data more effectively.

[0102] In some specific embodiments, each SA layer includes three steps:

[0103] 1. Sampling:

[0104] Use farthest point sampling to select a set of center points from the point cloud;

[0105] Farthest point sampling ensures that the sampled points are evenly distributed and can cover the entire point cloud.

[0106] 2. Grouping:

[0107] Taking each center point as a reference, select its K nearest neighbor points to construct a local area.

[0108] Each local area contains a center point and its neighbor points.

[0109] 3. Feature extraction:

[0110] Use a PointNet module to extract features from each local area.

[0111] Among them, the structure of the PointNet includes:

[0112] MLP (Multi-Layer Perceptron): Perform non-linear transformation on the features of each point.

[0113] Max Pooling (Global Feature Extraction Module): Perform max pooling on the point features within the local area to extract global features.

[0114] Finally, output the feature vector of each local area.

[0115] (6) The global features will be processed through an MLP layer to output the class probability distribution of the object, so as to achieve the purpose of classification construction.

[0116] (7) Use the loss function of cross-entropy loss to train the model by minimizing the difference between the predicted class and the true class.

[0117] (8) After the training stage, evaluate its accuracy through the validation set, check its classification accuracy, ensure that it is not overfitting, and obtain a relatively excellent recognition ability for typical bridge components (such as slab beams, bridge piers, etc.), such as Figure 3, the bridge component recognition model after being trained in this embodiment has a recognition accuracy greater than 90%.

[0118] In some specific embodiments, the working process of the recognition model of bridge components based on PointNet++ is as Figure 4 shown;

[0119] S102. Segmentation of bridge components based on region growing;

[0120] S1021. Input the point cloud data of bridge components obtained by the model recognition after being trained by S1;

[0121] S1022. Randomly select multiple seed points as the starting points (initial seed points) of the segmentation, and use the consistency of the normal vectors as its similarity criterion;

[0122] In this embodiment, the seed points are the starting points of region growing, which determine the initial position and the final result of the segmentation process.

[0123] S1023. Starting from the initial seed points, gradually check whether the adjacent unlabeled points meet the similarity criterion. If the criterion is met, add the point to the current region and continue to expand. The expansion process is usually recursive until no eligible points can be added, and a preliminary segmentation region is obtained.

[0124] S1024. When the expansion of a region is completed, select a new seed point to continue the segmentation. The algorithm terminates until the preset segmentation goals (such as the number of regions, size, etc.) are reached.

[0125] S1025. As Figure 5 , finally, the point cloud is divided into several non-overlapping regions or categories, and the category to which each point belongs is output, so as to achieve the purpose of separating the surface regions of bridge components.

[0126] S2. Establish a hybrid model of "point cloud + wireframe"

[0127] The standardized design and mass production characteristics of segment precast and assembled bridges determine that the point cloud data processing flow has significant reusability and patterned features. During the virtual assembly process of adjacent segment point cloud data, two core issues need to be addressed: optimizing the selection of the registration area and improving the iterative calculation efficiency. Therefore, when completing the necessary link of converting point cloud data into a parametric model, the data processing and iterative calculation amount should also be simplified as much as possible. This embodiment proposes a hybrid modeling method based on feature fusion: First, the geometric boundaries and skeleton structures of components are segmented from the discrete point cloud through a multi-scale feature extraction algorithm; the high-precision geometric information of the point cloud is fused with the wireframe model to form a "point cloud + wireframe" dual-mode fusion representation system containing rich geometric information. While ensuring the accurate expression of the outer contour of bridge components, fine-grained modeling is carried out on the key parts during the docking and assembly process, which helps to improve the assembly quality of bridge components during the actual construction process.

[0128] S201. Use the Canny edge detection algorithm to extract the geometric boundary information of the segmented component parts.

[0129] S202. Generate a skeleton from the extracted boundary information according to the thinning algorithm, and this skeleton can represent the geometric center and skeleton lines of the component.

[0130] In some specific embodiments, the specific process of the thinning algorithm is as follows:

[0131] 1. Input and initialization

[0132] Input a binary image or boundary information, where the pixel value of the target area (component) is 1 (representing the foreground), and the pixel value of the background area is 0. On this basis, the initialization process marks all foreground pixels as points to be processed and prepares to enter the thinning stage.

[0133] 2. Iterative thinning

[0134] During the iterative thinning process, the target area is gradually thinned by repeatedly removing boundary points until no more points can be removed. First, traverse all foreground pixels and check whether each pixel meets the deletion condition according to its 8-neighborhood or 4-neighborhood relationship to ensure that the deletion operation does not damage the connectivity of the target or change its topological structure. The deletion conditions include the simple point condition and endpoint protection: the simple point condition ensures that deleting this point does not affect the overall topological properties of the target; the endpoint protection mechanism avoids deleting pixels with only one adjacent point to maintain the integrity of the skeleton ends.

[0135] Next, for all pixels marked as deletable, remove them from the foreground and reset their pixel values to 0 (i.e., the background). After that, check if any pixels were deleted in the current iteration. If no pixels were deleted, it is considered that the termination condition has been reached and the algorithm stops running. Otherwise, if pixels were deleted, continue with the next iteration.

[0136] 3. Output Results

[0137] The final output result is a target skeleton composed of a series of single-pixel-wide lines. These lines represent the geometric centerline of the original target area, accurately reflecting the shape characteristics and internal structure of the target.

[0138] S203. Further, based on the generated skeleton information and combined with the point cloud data of bridge components, perform refined modeling on the key parts during the assembly process to ensure the improvement of the assembly quality of bridge components during the actual construction process.

[0139] The process of refined modeling is as follows:

[0140] 1. Based on the generated skeleton information and combined with the point cloud data of bridge components, mark the key point coordinates at the branch points and endpoints of the components, and extract their normal vectors and curvatures for subsequent modeling;

[0141] 2. Register and align the skeleton information with the point cloud data to ensure that both are in the same coordinate system, and extract the local point cloud data of the key points;

[0142] 3. Modeling steps:

[0143] 3.1 Preprocess the local point cloud data to remove noise and outliers;

[0144] 3.2 Fit a surface to the local point cloud to generate a refined geometric model, and generate a continuous surface based on the point cloud normal vector to achieve high precision;

[0145] 3.3 Add detailed information to the key parts (such as seams, connection points) of the refined geometric model, extract boundary points from the point cloud, and enhance the geometric representation at the seams;

[0146] 3.4 Check if the refined geometric model is consistent with the original point cloud.

[0147] Use the distance error from the point to the model (Hausdorff distance) to evaluate the accuracy of the refined geometric model. If it is found that the accuracy of some parts is insufficient, re-extract the local point cloud and repeat the above steps;

[0148] 3.5 Output the specific parameters (dimensions, angles) of the 3D model.

[0149] S204. Smooth and simplify the refined geometric model simultaneously, streamline unnecessary point cloud data according to the skeleton information, so that the skeleton information can be better combined with the point cloud data, achieving the purpose of ensuring the accuracy of the assembly process and effectively reducing the computational burden.

[0150] In some possible implementation manners, the bilateral filtering smoothing process is performed on the three-dimensional model by using the Open3D library in Python.

[0151] S205. The obtained model contains both the detailed data of the geometric edges of the components characterized by the skeleton information and the point cloud data refined based on the wireframe model, thus realizing the establishment of the "point cloud + wireframe" model.

[0152] S3. High-precision detection of the dimensions of bridge precast components

[0153] The dimension detection of bridge precast components is crucial for the assembly quality. Especially for strict dimension control of the connection parts, high stress concentration phenomena can be avoided, and potential weak links in the structure after assembly can be reduced. Aiming at the problems that the existing traditional manual methods are difficult in detecting the linear type of prestressed ducts in large precast components, with high labor intensity and long time consumption, a method for detecting the geometric features of bridge components based on the "point cloud + wireframe" hybrid model is proposed. In this implementation manner, the Laplacian eigenmap method is used to simplify complex three-dimensional data into two-dimensional images that are easy to process, and the Alpha shape algorithm is combined to accurately extract the edge features of the components. By calculating the distance from each pixel to the nearest edge based on the distance transformation technology, a distance map is generated, thereby realizing the high-precision measurement of the component dimensions. It effectively overcomes the limitations of traditional measurement technologies under complex geometric conditions and provides a new technical path for the precise detection of bridge precast components.

[0154] S301. Use the Laplacian eigenmap method to reduce the dimension of the three-dimensional data of the dual-mode fusion representation model containing point cloud and wireframe.

[0155] The Laplacian eigenmap method is an algorithm for dimensionality reduction and feature learning, which is widely used in the representation learning of non-linear data. It maps high-dimensional data to a low-dimensional space by preserving the local structure of the data. In the processing of point cloud data and geometric feature detection, the Laplacian eigenmap can help simplify complex three-dimensional data and retain its local geometric structure.

[0156] S3011. Mapping of three-dimensional point cloud data to two-dimensional images: Use the Laplacian eigenmap method to map the local structure information in the three-dimensional point cloud data to the two-dimensional space. In this implementation manner, by calculating the Laplacian operator of the point cloud, the global geometric structure of the data is retained while the dimension is reduced, making the subsequent image processing more convenient.

[0157] S3012. Feature mapping process: In Laplacian eigenmaps, a graph is first constructed to represent the local relationships in the point cloud data, and then the eigenvectors of the Laplacian matrix of the graph are calculated to obtain a low-dimensional representation. Finally, the obtained two-dimensional feature map can simplify the processing process while maintaining the original data structure.

[0158] S302. Alpha shape algorithm

[0159] The Alpha shape algorithm is an algorithm for describing the shape of a point set. It controls the complexity of the point set by defining a parameter (Alpha value) and can effectively extract the boundaries and outlines in the point cloud. In this embodiment, the Alpha shape algorithm is used to extract the outer contours and edge features of the component, providing accurate edge information for subsequent dimension measurement.

[0160] S3021. Edge extraction: Based on the Alpha shape algorithm, extract the edge lines of the component. The edges of the Alpha shape are computationally relatively simple and can be flexibly adjusted according to the set Alpha value to ensure the accuracy of the edge information.

[0161] S3022. Edge optimization and connection: In the extracted edge data, there may be some noise or incomplete connections. It is necessary to smooth the edges and repair the missing edge connections through an adjacent point connection algorithm (such as the RANSAC algorithm) to ensure the integrity and continuity of the edge model.

[0162] S303. Dimension measurement

[0163] S3031. Calculate the distance from each point in the point cloud data to the nearest edge. After the geometric boundary and skeleton information of the component are extracted, the distance transformation algorithm can accurately calculate the distance from each pixel point (or each point in the point cloud) in the point cloud data to the edge. Specifically:

[0164] 1. Input the point cloud data (including skeleton information and geometric boundary information);

[0165] 2. Use the edge information extracted in step S302;

[0166] 3. Use the KD-Tree data structure to accelerate the calculation of the distance from each point in the point cloud to the edge point.

[0167] S3032. By calculating the distance from each pixel point to the nearest edge, a distance map is generated. The distance map in this embodiment can not only help find the relative position of the point to the edge in space but also support the dimension measurement of the component.

[0168] S3033. Based on the generation of the distance map, precise measurement of the component dimensions can be carried out. Especially during the component assembly process, the accuracy of the joint parts can be adjusted according to the measurement results to ensure the accuracy and quality of the docking of each segment.

[0169] Each embodiment in this specification is described in a related manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment.

[0170] The above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are all included in the protection scope of the present invention.

Claims

1. A method for determining the geometric features of bridge components based on a hybrid model, characterized in that The specific steps are as follows: S1. Construct a PointNet++-based bridge component recognition model and train the model, and then use the model to identify and segment the bridge component point cloud data; The bridge component recognition model includes a relative position encoding module and a spatial attention module; The relative position encoding module includes an input module, a feature extraction module, a global feature extraction module, a normalization module, an embedding module, a feature preprocessing module, and an output module connected in sequence; The input module includes a point feature matrix of N×d, where N is the number of points in the bridge component point cloud and d is the feature dimension of each point; The feature extraction module includes three SA layers, and each SA layer includes a PointNet module. The PointNet module includes an MLP and a global feature extraction module; the MLP is used to perform a non-linear transformation on the features of each point, and the global feature extraction module is used to perform max pooling on the point features in the local area to extract global features; The normalization module normalizes the features; The embedding module uses the normalized features as an embedding representation and encodes the relative position information into a vector of a fixed dimension; The feature preprocessing module preprocesses the embedded features through two layers of MLP; The output module includes a final feature matrix of N×d; The spatial attention module includes a self-attention mechanism for implementing the calculation process of attention weights; The training process of the bridge component recognition model is specifically as follows: (1) Process the bridge component point cloud data in batches. Each batch contains multiple points, and each point contains coordinate positions (x, y, z) and additional feature values for each point; The additional feature values include normal vectors, surface directions, curvatures, and concavities and convexities; (2) Encode the batch-processed bridge component point cloud data; (3) Standardize the encoded bridge component point cloud data; (4) Perform augmentation processing on the bridge component point cloud data, including rotation, flipping, and scaling; (5) Insert a spatial attention module between the first and second SA layers of the bridge component recognition model; input the augmented point cloud data into the bridge component recognition model to extract local features and generate a global feature vector representing the entire point cloud, The execution of each SA layer includes three steps: a. Sampling: Use farthest point sampling to select a set of center points from the point cloud; b. Grouping: Based on each center point, select its nearest K neighbor points to construct a local area. Each local area contains a center point and its neighbor points; c. Feature extraction: Use a PointNet module to extract features for each local area; (6) The global feature vector is further processed through an MLP layer to output the class probability distribution of the object; (7) Use a loss function of cross-entropy loss to train the model by minimizing the difference between the predicted class and the true class; S2. Extract the geometric boundaries and skeletons by multi-scale feature extraction, and fuse the point cloud with the wireframe model to construct a dual-mode fusion representation model including the point cloud and the wireframe; S3. Determine the dimensions of the bridge components.

2. The method for determining the geometric characteristics of bridge components based on a hybrid model according to claim 1, wherein The self-attention mechanism is specifically as follows: The input feature map is X ∈ R H×W×C , where H is the height, W is the width, C is the number of channels, and R represents the set of real numbers; First, by performing a linear transformation on the input feature map, query vector Q, key vector K, and value vector V are calculated: Q = XW Q K = XW K V = XW V Among them, W Q , W K , W V ∈R C×d is the learned weight matrix; d is the feature dimension of each point; The attention weights are calculated. First, the similarity score A between Q and K is calculated: Among them, is the result of the scaled dot product; the softmax function is used to normalize the scores; T is the transpose symbol; Finally, the attention weights A are applied to the value vector V to obtain the final output Y: Y = AV.

3. A method for determining the geometric features of bridge components based on a hybrid model according to claim 1, characterized in that The specific process of segmenting the bridge components is as follows: Randomly select multiple seed points in the point cloud data of the bridge components identified by the bridge component recognition model as the starting points for segmentation; Starting from the initial seed points, gradually check whether the adjacent unlabeled points meet the normal consistency similarity criterion; If the criterion is met, add the point to the current region and continue to expand; until no eligible points can be added, a preliminary segmentation region is obtained; After the expansion of a region is completed, select a new seed point to continue the segmentation until the preset segmentation target is reached; finally, the point cloud is divided into several non-overlapping regions or categories, and the category to which each point belongs is output to complete the segmentation of the bridge components.

4. A method for determining geometric features of bridge components based on a hybrid model according to claim 1, characterized in that, The specific process of constructing the dual-mode fusion representation model of the point cloud and wireframe is as follows: S201. Extract the geometric boundary information of the segmented component parts; S202. Generate a skeleton based on the extracted boundary information. The skeleton includes the geometric center of the component and the skeleton lines; S203. According to the generated skeleton information combined with the point cloud data of the bridge components, perform refined modeling on the key parts during the assembly process; S204. Smooth and simplify the refined modeling results; obtain a dual-mode fusion representation model containing point clouds and wireframes.

5. The method for determining the geometric features of bridge components based on a hybrid model according to claim 4, wherein The specific process of the refined modeling in S203 is as follows: Based on the generated skeleton information combined with the point cloud data of the bridge components, mark the key point coordinates at the branch points and end points of the components, and extract their normal vectors and curvatures; then, register and align the skeleton information with the point cloud data to ensure that the two are in the same coordinate system, and then extract the local point cloud data of the key points; preprocess the local point cloud data to remove noise and outliers; perform surface fitting on the local point cloud to generate a refined geometric model, and generate a continuous surface based on the point cloud normal vector; add detailed information at the key parts, and extract boundary points from the point cloud to strengthen the geometric performance at the seams; check whether the refined geometric model is consistent with the original point cloud, and use the distance error from the point to the model to evaluate the accuracy of the refined geometric model. If the accuracy is found to be insufficient, re-extract the local point cloud and repeat the above steps; finally, output the specific parameters of the refined geometric model.

6. The method for determining the geometric features of bridge components based on a hybrid model according to claim 1, wherein The specific process of S3 is as follows: S301. Map the three-dimensional point cloud data of the dual-mode fusion representation model containing point clouds and wireframes to two dimensions; specifically, first construct a graph to represent the local relationships in the three-dimensional point cloud data of the dual-mode fusion representation model containing point clouds and wireframes, and then obtain the low-dimensional representation by calculating the eigenvectors of the Laplacian matrix of the graph; S302. Extract the edge information of the bridge components; S303. Determine the size of the bridge components.

7. A method for determining the geometric characteristics of bridge components based on a hybrid model according to claim 6, characterized in that The specific process of S302 is as follows: Extract the edge lines of the component based on the Alpha shape algorithm, and flexibly adjust the extraction accuracy by setting different Alpha values; Smooth the extracted edge data, and use the adjacent point connection algorithm to repair the missing edge connections to ensure the integrity and continuity of the edge model, and finally obtain an optimized edge model.

8. A method for determining geometric features of bridge components based on a hybrid model according to claim 6, characterized in that The specific process of S303 is as follows: Use the KD-Tree data structure to accelerate the calculation of the distance from each point in the point cloud to the nearest edge, generate a distance map, and accurately measure the size of the component based on the distance map.

Images