Hybrid model-based bridge member geometric feature determination method

Abstract

The invention provides a method for determining geometric features of a bridge member based on a hybrid model, which comprises the following steps: firstly, constructing and training a bridge member identification model, and then identifying and segmenting point cloud data of the bridge member by using the model. Thirdly, segmenting geometric boundaries and skeletons through multi-scale feature extraction, fusing the geometric boundaries and skeletons with the wireframe model, and constructing a dual-mode fusion representation model containing point clouds and wireframes; and finally, based on the model and the point cloud data, calculating the distance from each point to the nearest edge to generate a distance map, and performing accurate measurement on the size of the bridge member by using the distance map to ensure the butt joint precision of each section. According to the method, region growth segmentation and a'point cloud + wireframe 'mixed model are used, data processing is simplified, and calculation efficiency is improved. High-precision size measurement is realized by utilizing Laplace feature mapping and an Alpha shape algorithm, and the method is suitable for bridge component detection under complex geometric conditions.

Description

Technical Field

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

[0002] Segmental prefabricated components are an important part of the industrialized construction of bridges. The actual quality of the components after they are produced in the factory has a crucial impact on the on-site construction. Segmental prefabricated components are usually assembled on site, which has higher requirements for the linear control and splicing error of the assembled segments. Any segment assembly error will lead to quality and efficiency problems in the final bridge. In order to ensure that the quality of segmental prefabricated components meets the standards and reduce construction assembly errors and construction risks, spatially related prefabricated components usually need to be temporarily assembled before leaving the factory, that is, pre-assembled, so as to provide guidance on the processing and linear control of the components before actual assembly. Pre-assembly requires the addition of different temporary lifting equipment and safety protection measures. Although this method is closer to reality, it has disadvantages such as large safety hazards, long pre-assembly period, large site occupation, high tire frame and labor costs. In addition, for complex environments such as mountainous areas and across (rivers) and seas, the pre-assembly of (extra) large-span bridge components cannot be fully realized due to limitations in site and equipment conditions. Unit modularization is often used for pre-assembly, which increases project costs, delays construction progress, and to some extent affects the quality and efficiency of actual assembly.

[0003] With the demand for the use of segmental prefabricated components in industrialized bridge construction and the continuous development of science and technology, virtual assembly theories and methods with high efficiency and quality of bridge component assembly have gradually shown application prospects in bridge construction.

[0004] However, there are still some deficiencies in the research on the technical reserve for realizing the automatic identification of prefabricated bridge components and obtaining their finite element models: Traditional methods have some limitations in bridge component recognition and segmentation. Because they rely on artificially designed features, they cannot fully capture complex shapes and details, resulting in low recognition accuracy. Insufficient use 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, and have low robustness; they tend to ignore the contextual relationship between components, which can easily lead to misrecognition or omission.

[0005] In terms of point cloud data processing, the simulated assembly and matching process of adjacent segment point cloud data involves the selection of alignment areas and a large number of iterative calculations, which can easily lead to unnecessary large data parameters. It is impossible to ensure the accurate expression of the outer contour of the bridge components while performing fine modeling of the key parts in the docking assembly process.

[0006] In the inspection of the size of prefabricated bridge components, the existing traditional manual methods have great difficulties in large prefabricated components, especially in the detection of prestressed pipe line shape. Because prestressed pipes are usually buried inside or on the surface of components and have complex line shapes, manual inspection has low accuracy and is prone to errors. In addition, manual inspection requires a lot of time and energy, and is labor-intensive. Especially in large components, the inspection process is time-consuming and inefficient. As the scale and complexity of components increase, the limitations of traditional methods become more obvious, and more efficient and intelligent technical means are urgently needed to improve inspection accuracy and efficiency. Summary of the invention

[0007] 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 cost, low efficiency and limited site in the pre-assembly process of traditional prefabricated bridge components, as well as the technical difficulties of insufficient component recognition accuracy, low efficiency in point cloud data processing, and difficult manual inspection of key parts such as prestressed pipes in virtual assembly technology.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is a method for determining geometric features of bridge components based on a hybrid model, which is specifically carried out in the following steps: S1. Build and train a bridge component recognition model, and then use the model to recognize and segment bridge component point cloud data; S2, extract and segment the geometric boundary and skeleton through multi-scale features, and fuse them with the wireframe model to build a dual-mode fusion representation model including point cloud and wireframe; S3. Determine the size of bridge components.

[0009] Furthermore, the bridge component recognition model described in S1 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; An input module includes an N×d point feature matrix, where N is the number of points in the bridge component point cloud and d is the feature dimension of each point; A feature extraction module includes three SA layers, each of which includes a PointNet module, wherein the PointNet module includes an MLP and a global feature extraction module; the MLP is used to perform nonlinear transformation on the features of each point, and the global feature extraction module is used to perform maximum pooling on the point features in the local area to extract global features; Normalization module, normalizes the features; The embedding module uses the normalized features as embedding representation and encodes the relative position information into a vector of fixed dimension; Feature preprocessing module, which preprocesses the embedded features through two layers of MLP; Output module, including the final feature matrix of N×d; The spatial attention module includes a self-attention mechanism for implementing the calculation process of attention weights.

[0010] Furthermore, the self-attention mechanism is specifically as follows: The input feature map is ,in is the height, is the width, is the number of channels, represents the set of real numbers; First, the query vector is calculated by linearly transforming the input feature map , key vector Sum value vector : ; in, is the learned weight matrix; is the feature dimension of each point; To calculate the attention weight, first calculate and The similarity score between : ; in, is the result of scaling the dot product; The function is used to normalize the score; T is the transpose symbol; Finally, the attention weight Applied to a vector of values , and get the final output : .

[0011] Furthermore, the training process of the bridge component recognition model is specifically as follows: (1) Processing bridge component point cloud data in batches, each batch containing multiple points, each point containing coordinate position (x, y, z) and additional feature values ​​for each point; the additional feature values ​​include normal vector, surface direction, curvature, and concavity; (2) Encoding the batch-processed bridge component point cloud data; (3) Standardize the encoded bridge component point cloud data; (4) Enhance the point cloud data of bridge components, including rotation, flipping, and scaling; (5) Inserting a spatial attention module between the first and second SA layers of the bridge component recognition model; inputting the enhanced 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 consists of three steps: a. Sampling: Use farthest point sampling to select a set of center points from the point cloud; b. Grouping: Taking each center point as the benchmark, select its nearest K neighboring points to construct local areas. Each local area contains a center point and its neighboring points. c. Feature extraction: Use a PointNet module to extract features from each local area; (6) The global feature vector is then processed through an MLP layer to output the class probability distribution of the object; (7) Use the cross entropy loss function to train the model by minimizing the difference between the predicted category and the true category.

[0012] Furthermore, the segmentation process of the bridge component is specifically as follows: randomly selecting multiple seed points as the starting points of segmentation in the bridge component point cloud data after the bridge component recognition model is identified; starting from the initial seed point, gradually checking whether the adjacent unmarked points meet the consistency similarity criterion of the normal line; if the criterion is met, adding the point to the current area and continuing to expand; until there are no points that meet the conditions to be added, a preliminary segmentation area is obtained; When the expansion of an area is completed, a new seed point is selected to continue segmentation until the preset segmentation target is reached; finally, the point cloud is divided into several non-overlapping areas or categories, and the category to which each point belongs is output to complete the segmentation of bridge components.

[0013] Furthermore, the construction process of the dual-mode fusion representation model of point cloud and wireframe is specifically as follows: S201, extracting geometric boundary information of the segmented component parts; S202, generating a skeleton based on the extracted boundary information, wherein the skeleton includes a geometric center and a skeleton line of the component; S203, performing detailed modeling of key parts in the assembly process based on the generated skeleton information and the bridge component point cloud data; S204, smoothing and simplifying the refined modeling results; obtaining a dual-mode fusion representation model including point cloud and wireframe.

[0014] Furthermore, the refined modeling process described in S203 is specifically as follows: based on the generated skeleton information combined with the point cloud data of the bridge component, the coordinates of the key points are marked at the branch points and end points of the component, and their normal vectors and curvatures are extracted; then, the skeleton information is aligned with the point cloud data to ensure that the two are in the same coordinate system, and then the local point cloud data of the key points is extracted; the local point cloud data is preprocessed to remove noise and outliers; the local point cloud is fitted with a surface to generate a refined geometric model, and a continuous surface is generated based on the point cloud normal vector; detailed information is added to the key parts, and boundary points are extracted from the point cloud to enhance the geometric performance of the joints; check whether the refined model is consistent with the original point cloud, and use the point-to-model distance error to evaluate the accuracy of the refined geometric model. If the accuracy is found to be insufficient, the local point cloud is re-extracted and the above steps are repeated; finally, the specific parameters of the three-dimensional model are output.

[0015] Furthermore, the specific process of S3 is as follows: S301, mapping the three-dimensional point cloud data of the dual-mode fusion representation model including the point cloud and the wireframe to two dimensions; specifically, first constructing a graph to represent the local relationship in the three-dimensional point cloud data of the dual-mode fusion representation model including the point cloud and the wireframe, and then obtaining a low-dimensional representation by calculating the eigenvector of the Laplacian matrix of the graph; S302, extracting edge information of bridge components; S303. Determine the size of bridge components.

[0016] Furthermore, the specific process of S302 is: extracting the edge lines of the component based on the Alpha shape algorithm, and flexibly adjusting the extraction accuracy by setting different Alpha values; smoothing the extracted edge data, and using 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.

[0017] Furthermore, the specific process of S303 is: using the KD-Tree data structure to accelerate the calculation of the distance from each point in the point cloud to the nearest edge, generating a distance map, and accurately measuring the size of the component based on the distance map.

[0018] Compared with the prior art, the beneficial effects of the present invention include the following points: 1. The present invention is a fast and reliable geometric feature detection algorithm. Compared with the low efficiency and long calculation time of general detection algorithms, it has the advantages of high precision, high recognition efficiency, accurate segmentation type, short time consumption, and accurate key dimensions.

[0019] 2. The efficient recognition and segmentation method used in the present invention is accurate and efficient. The improved bridge component recognition model based on PointNet++ combined with the spatial attention module and position encoding can capture complex shapes and details. The model obtained through the effective training set and test set has high credibility. The segmentation model algorithm based on regional growth can effectively extract the feature information of the component, make full use of the spatial information to effectively process the relative position between the components, and achieve the purpose of surface separation based on algorithm learning.

[0020] 3. The "point cloud + wireframe" hybrid model proposed in the present invention has high applicability. Compared with the traditional point cloud model, which has unnecessarily large data and slow calculation and low efficiency, it simplifies data processing and iterative calculation when completing the necessary link of converting point cloud data into a parametric model. The geometric boundary and skeleton information of the segmented component parts are extracted through the wireframe model, and the high-precision geometric information of the point cloud is integrated with the wireframe model.

[0021] 4. The high-precision detection method for the size of prefabricated bridge components proposed in the present invention is based on the "point cloud + wireframe" model to establish an algorithm, uses the Laplace feature mapping method to simplify complex three-dimensional data into an easy-to-process two-dimensional image, combines the Alpha shape algorithm to accurately extract the edge features of the component, and then calculates the distance from each pixel to the nearest edge based on the distance transformation technology to achieve high-precision measurement of the component size. It can overcome the detection problem of bridge components under complex geometric conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 It is the technical roadmap of this implementation method; Figure 2 is a schematic diagram of a relative position encoding module; Figure 3 is a recognition model recognition effect diagram of the bridge component in this implementation mode; Figure 4 This is the identification model identification process of the bridge component in this implementation mode; Figure 5 This is the segmentation result based on region growing in this embodiment. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] Existing segmental prefabricated components are usually constructed by on-site assembly, which has higher requirements on the linear control and splicing error of the assembled segments. Any segment assembly error will lead to quality and efficiency problems in the final bridge.

[0026] This implementation method proposes a method for determining the geometric features of bridge components based on a hybrid model. First, PointNet++ is used as the basic architecture to extract the multi-scale geometric features of bridge components. The extracted features are processed using a segmentation model algorithm based on region growing to achieve accurate segmentation of the key parts of the components. Based on the point cloud data of the bridge components, when the necessary steps of converting the point cloud data into a parametric model are completed, the high-precision geometric information of the point cloud is fused with the wireframe model by simplifying data processing and reducing the amount of iterative calculations to form a 'point cloud + wireframe' hybrid model containing rich geometric information. Finally, a "point cloud + wireframe" hybrid model bridge component geometric feature detection algorithm is proposed to achieve high-precision measurement of component dimensions. It can overcome the limitations of traditional measurement technology under complex geometric conditions and provide a new technical path for the accurate detection of prefabricated bridge components.

[0027] In some specific embodiments, such as Figure 1 The hybrid model-based method for determining geometric features of bridge components mainly includes efficient recognition and segmentation of bridge components, construction of a "point cloud + wireframe" hybrid model, and high-precision detection of bridge component sizes.

[0028] S1. Efficient identification and segmentation of bridge components S101, constructing and training a bridge component recognition model, and using the recognition model to recognize the collected point cloud data; PointNet++ is a deep learning architecture for point cloud data processing, mainly used in 3D object recognition, classification, segmentation and other tasks. Unlike traditional convolutional neural networks (CNNs), PointNet directly processes raw point cloud data that has not been gridded or rasterized, avoiding the risk of point cloud data loss. Unlike the original PointNet, PointNet++ introduces a hierarchical structure to enhance the understanding of local details, allowing it to more effectively capture the local features of point clouds. This implementation is based on the traditional PointNet method, combining the spatial attention mechanism (SAM) module with the relative position encoding module to improve the accuracy of bridge component feature extraction and segmentation.

[0029] The bridge component recognition model based on PointNet++ in this embodiment includes a relative position encoding module and a spatial attention module; In some specific implementations, 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 into the neural network as initial features. Relative position encoding is achieved by encoding the relative position coordinates between a point and its neighboring points: ; in, Representative Points The position encoding vector, , Respectively represent points and neighbors The center point coordinates, Represents the encoding function, which is used to Converted into an encoded form to help the model aggregate local features and understand contextual information.

[0030] Before encoding, the K nearest neighbor algorithm is used to convert the points Nearby points are divided into their neighboring points. In this embodiment, the encoding process can be regarded as building a PointNet module, extracting relative position information through three layers of MLP and maximum pooling. Then point convolution is used to match the dimensions. Subsequently, the encoding is normalized and embedded or added to the neural network as the initial feature, followed by feature preprocessing using two layers of MLP for aggregation.

[0031] In some specific embodiments, the relative position encoding module structure is as follows: Figure 2 As shown, it 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 which are connected in sequence.

[0032] In some possible implementations, the input module includes an N×d point feature matrix, where N is the number of points in the bridge component point cloud, and d is the feature dimension (coordinate, normal vector) of each point; the feature extraction module performs nonlinear transformation on the input features through a three-layer MLP (multi-layer perceptron) to extract high-level features of the points; the global feature extraction module performs maximum pooling on the features to obtain global features (relative position information); the normalization module normalizes the features to improve the stability and convergence speed of model training; the embedding module uses the normalized features as embedded representations and encodes the relative position information into a vector of fixed dimension; the feature preprocessing module preprocesses the embedded features through a two-layer MLP; the output module includes an N×d final feature matrix for subsequent output tasks.

[0033] In some specific embodiments, the spatial attention mechanism module includes a self-attention mechanism for implementing the attention weight calculation process; the self-attention mechanism can adaptively learn the dependency between each position based on the input feature map, thereby achieving weighted importance of different positions.

[0034] In some specific implementations, the self-attention mechanism is as follows: The input feature map is ,in is the height, is the width, is the number of channels. First, by performing a linear transformation on the input feature map, we calculate and , the formulas are shown in formula (1) to formula (3).

[0035] (1) (2) (3) Where: is the learned weight matrix; is the feature dimension of each point.

[0036] To calculate the attention weight, first calculate and The similarity score between them is calculated using dot-product attention. ; Where: is the result of scaling the dot product; Function used to normalize the scores.

[0037] Finally, the attention weights are applied to , the output formula of weighted sum is shown in (4).

[0038] (4) Through the above formula, the spatial attention module can calculate the attention weight of each position according to the input feature map, thereby realizing weighted fusion of feature representations at different positions and improving the model's ability to focus on local features.

[0039] (1) The collected point cloud data of bridge components are processed in batches. Each batch contains multiple points, and each point contains the coordinate position (x, y, z) and its additional feature values ​​(including normal vector, surface direction, curvature, convexity, etc.).

[0040] (2) Encode the data according to the relative position encoding module before inputting the data.

[0041] (3) Standardize the point cloud processed in (2) by first translating the point cloud to the origin and then scaling it to a uniform size range (such as a unit cube).

[0042] (4) Data enhancement processing: processing point cloud data before training by rotating, flipping, scaling, etc. to improve the robustness of the model.

[0043] (5) The input point cloud data will pass through a multi-layer perceptron (MLP) to extract local features from each point, and then the point cloud will be processed step by step through a hierarchical structure. At each layer, local point sets are first clustered or neighborhood selection (such as k-nearest neighbors or sphere sampling) is performed, and then local features are extracted through a local network. The point features are then aggregated through maximum pooling to generate a global feature vector representing the entire point cloud.

[0044] In some specific embodiments, the model based on PointNet++ of the present application also includes three SA layers, which gradually downsample and extract features; the point cloud data enters an attention weight processing module before entering the first SA layer and before entering the second SA, and can adaptively adjust the feature weights according to the importance of different positions in the data, thereby processing the point cloud data more efficiently.

[0045] In some specific implementations, each SA layer includes three steps: 1. Sampling: Select a set of center points from the point cloud using farthest point sampling; The farthest point sampling ensures that the sampling points are evenly distributed and can cover the entire point cloud.

[0046] 2. Grouping: Taking each center point as the benchmark, select its nearest K neighbor points to construct the local area.

[0047] Each local region consists of a center point and its neighboring points.

[0048] 3. Feature extraction: A PointNet module is used to extract features from each local area.

[0049] The structure of PointNet includes: MLP (Multi-layer Perceptron): Performs nonlinear transformation on the features of each point.

[0050] Max Pooling (global feature extraction module): Perform maximum pooling on point features in local areas to extract global features.

[0051] Finally, the feature vector of each local area is output.

[0052] (6) The global features will be processed through an MLP layer to output the class probability distribution of the object to achieve the purpose of classification construction.

[0053] (7) Use the cross entropy loss function to train the model by minimizing the difference between the predicted category and the true category.

[0054] (8) After the training phase, the accuracy is evaluated through the validation set to check the classification accuracy and ensure that it is not overfitting, so as to obtain a good recognition ability for typical bridge components (plate beams, piers, etc.). Figure 3 , the bridge component recognition model after training in this implementation mode has a recognition accuracy greater than 90%.

[0055] In some specific implementations, the workflow of the bridge component recognition model based on PointNet++ is as follows: Figure 4 As shown; S102. Bridge component segmentation based on region growing; S1021, inputting the bridge component point cloud data obtained by model recognition after training in S1; S1022, randomly selecting multiple seed points as the starting points of segmentation (initial seed points), and using the consistency of normals as their similarity criterion; In this implementation, the seed point is the starting point of region growing and determines the initial position and final result of the segmentation process.

[0056] S1023, starting from the initial seed point, check whether the adjacent unmarked points meet the similarity criteria step by step. If the criteria are met, add the point to the current region and continue to expand. The expansion process is usually recursive until there are no points that meet the conditions to be added, and the preliminary segmented region is obtained.

[0057] S1024: When the expansion of a region is completed, a new seed point is selected to continue segmentation until the preset segmentation target (such as the number and size of regions) is reached, and the algorithm terminates.

[0058] S1025, such as Figure 5 ,Finally, the point cloud is divided into several non-overlapping areas or categories, and the category to which each point belongs is output, so as to achieve the purpose of separating the surface areas of bridge components.

[0059] S2. Build a hybrid model of "point cloud + wireframe" The standardized design and mass production characteristics of segmental prefabricated assembled bridges determine that their point cloud data processing process has significant reusability and pattern characteristics. In the virtual assembly process of adjacent segment point cloud data, it is necessary to focus on solving two core problems: optimal selection of registration areas and improvement of iterative calculation efficiency. Therefore, when completing the necessary steps of converting point cloud data into a parametric model, data processing and iterative calculations should be simplified as much as possible. This implementation method proposes a hybrid modeling method based on feature fusion: first, the geometric boundaries and skeleton structures of the 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 the bridge component, the key parts of the docking assembly process are refined modeling, which helps to improve the assembly quality of bridge components during actual construction.

[0060] S201, using the Canny edge detection algorithm to extract geometric boundary information of the segmented component parts.

[0061] S202: Generate a skeleton from the extracted boundary information according to a refinement algorithm, where the skeleton can represent the geometric center and skeleton line of the component.

[0062] In some specific implementations, the specific process of the refinement algorithm is: 1. Input and initialization Input binary image or boundary information, where the pixel value of the target area (component) is 1 (indicating foreground) and the pixel value of the background area is 0. Based on this, the initialization process marks all foreground pixels as points to be processed and prepares to enter the refinement stage.

[0063] 2. Iterative refinement During the iterative refinement process, the target area is gradually refined by cyclically removing boundary points until no more points can be removed. First, all foreground pixels are traversed and each pixel is checked to see if it meets the deletion condition based on its 8-neighborhood or 4-neighborhood relationship to ensure that the deletion operation does not destroy the connectivity of the target or change its topological structure. The deletion conditions include simple point conditions and endpoint protection: the simple point condition ensures that deleting the point will not affect the overall topological properties of the target; the endpoint protection mechanism prevents pixels with only one neighboring point from being deleted to maintain the integrity of the skeleton end.

[0064] Next, for all pixels marked as deletable, remove them from the foreground and reset their pixel values ​​to 0 (i.e., background). After that, check whether any pixels are deleted in the current iteration. If no pixels are deleted, it is considered that the termination condition is met and the algorithm stops running. On the contrary, if any pixels are deleted, continue to the next round of iteration.

[0065] 3. Output results The final output is a target skeleton composed of a series of single-pixel-width lines, which represent the geometric center lines of the original target area and accurately reflect the shape characteristics and internal structure of the target.

[0066] S203. Further, based on the generated skeleton information and the bridge component point cloud data, the key parts in the assembly process are finely modeled to ensure the improvement of the assembly quality of the bridge components in the actual construction process.

[0067] The detailed modeling process is as follows: 1. Based on the generated skeleton information and the point cloud data of the bridge components, mark the coordinates of the key points at the branch points and end points of the components, and extract their normal vectors and curvatures for subsequent modeling; 2. Register and align the skeleton information with the point cloud data to ensure that they are in the same coordinate system and extract the local point cloud data of key points; 3. Modeling steps: 3.1 Preprocess the local point cloud data to remove noise and outliers; 3.2 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 to achieve high accuracy; 3.3 Add detailed information to key parts (such as seams and connection points) of the refined geometric model, extract boundary points from the point cloud, and enhance the geometric representation of the seams; 3.4 Check whether the refined geometric model is consistent with the original point cloud.

[0068] Use the distance error (Hausdorff distance) from the point to the model 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; 3.5 Output the specific parameters (size, angle) of the 3D model.

[0069] S204. At the same time, the refined geometric model is smoothed and simplified, and unnecessary point cloud data is streamlined according to the skeleton information, so that the skeleton information is better combined with the point cloud data, so as to achieve the purpose of ensuring the accuracy of the assembly process and effectively reducing the calculation burden.

[0070] In some possible implementations, bilateral filtering and smoothing processing is performed on the three-dimensional model using the Open3D library in Python.

[0071] S205. The obtained model includes both detailed data of the component geometric edges represented by the skeleton information and point cloud data simplified based on the wireframe model, thereby realizing the establishment of a "point cloud + wireframe" model.

[0072] S3. High-precision detection of the size of prefabricated bridge components The dimensional inspection of prefabricated bridge components is crucial to the quality of assembly, especially strict dimensional control of the connection parts, which can avoid high stress concentration and reduce potential weak links in the assembled structure. In view of the problems that the existing traditional manual methods are difficult, labor-intensive, and time-consuming in the linear detection of prestressed pipes in large prefabricated components, a hybrid model of "point cloud + wireframe" bridge component geometric feature detection method is proposed. This implementation method uses the Laplace feature mapping method to simplify complex three-dimensional data into easy-to-process two-dimensional images, and combines the Alpha shape algorithm to accurately extract component edge features. By calculating the distance from each pixel to the nearest edge based on the distance transformation technology, a distance map is generated, thereby achieving high-precision measurement of component dimensions. It effectively overcomes the limitations of traditional measurement technology under complex geometric conditions and provides a new technical path for the accurate detection of prefabricated bridge components.

[0073] S301, using a Laplace eigenmapping method, reducing the dimension of the three-dimensional data of the dual-mode fusion representation model including the point cloud and the wireframe; The Laplace eigenmapping method is an algorithm for dimensionality reduction and feature learning. It is widely used in representation learning of nonlinear data. It maps high-dimensional data to low-dimensional space by maintaining the local structure of the data. In point cloud data processing and geometric feature detection, the Laplace eigenmapping can help simplify complex three-dimensional data and retain its local geometric structure.

[0074] S3011, mapping of 3D point cloud data to 2D image: using Laplace feature mapping method, local structure information in 3D point cloud data is mapped to 2D space. This embodiment retains the global geometric structure of the data by calculating the Laplace operator of the point cloud, while reducing the dimension, making subsequent image processing easier.

[0075] S3012, Feature Mapping Process: In Laplacian feature mapping, a graph is first constructed to represent the local relationship in the point cloud data, and then the eigenvector of the Laplacian matrix of the graph is 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.

[0076] S302, Alpha shape algorithm The Alpha shape algorithm is an algorithm used to describe 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 shapes in the point cloud. In this embodiment, the Alpha shape algorithm is used to extract the outer contour and edge features of the component to provide accurate edge information for subsequent dimensional measurement.

[0077] S3021, edge extraction: extract the edge lines of the component based on the Alpha shape algorithm. The edge of the Alpha shape is relatively simple to calculate, and the extraction accuracy can be flexibly adjusted according to the set Alpha value to ensure the accuracy of the edge information.

[0078] S3022, edge optimization and connection: There may be some noise or incomplete connections in the extracted edge data. The edges need to be smoothed and the missing edge connections need to be repaired through adjacent point connection algorithms (such as the RANSAC algorithm) to ensure the integrity and continuity of the edge model.

[0079] S303, Dimension Measurement S3031. Calculate the distance from each point in the point cloud data to the nearest edge. After the component geometric boundary and skeleton information are extracted, the distance transformation algorithm can accurately calculate the distance from each pixel in the point cloud data (or each point in the point cloud) to the edge. Specifically: 1. Input point cloud data (including skeleton information and geometric boundary information); 2. Use the edge information extracted in step S302; 3. Use KD-Tree data structure to accelerate the distance calculation from each point in the point cloud to the edge point.

[0080] S3032: Generate a distance map by calculating the distance between each pixel and the nearest edge. In this embodiment, the distance map can not only help find the relative position of a point to an edge in space, but also provide support for the size measurement of a component.

[0081] S3033. Based on the generation of the distance map, the component dimensions can be accurately measured. Especially during the component assembly process, the accuracy of the joints can be adjusted according to the measurement results to ensure the accuracy and quality of the docking of each segment.

[0082] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and 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 partial description of the method embodiment.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for determining geometric features of bridge components based on a hybrid model, characterized in that: Follow these steps: S1. Build a bridge component recognition model based on PointNet++ and train the model, and then use the model to recognize and segment bridge component point cloud data; S2, segmenting geometric boundaries and skeletons through multi-scale feature extraction, and fusing the point cloud with the wireframe model to build a dual-mode fusion representation model including point cloud and wireframe; S3. Determine the size of bridge components.

2. A method for determining geometric features of bridge components based on a hybrid model according to claim 1, characterized in that: The bridge component recognition model described in S1 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; An input module includes an N×d point feature matrix, where N is the number of points in the bridge component point cloud and d is the feature dimension of each point; A feature extraction module includes three SA layers, each of which includes a PointNet module, wherein the PointNet module includes an MLP and a global feature extraction module; the MLP is used to perform nonlinear transformation on the features of each point, and the global feature extraction module is used to perform maximum pooling on the point features in the local area to extract global features; Normalization module, normalizes the features; The embedding module uses the normalized features as embedding representation and encodes the relative position information into a vector of fixed dimension; Feature preprocessing module, which preprocesses the embedded features through two layers of MLP; Output module, including the final feature matrix of N×d; The spatial attention module includes a self-attention mechanism for implementing the calculation process of attention weights.

3. A method for determining geometric features of bridge components based on a hybrid model according to claim 2, characterized in that: The self-attention mechanism is specifically: The input feature map is ,in is the height, is the width, is the number of channels, represents a set of real numbers; first, the query vector is calculated by linearly transforming the input feature map , key vector Sum value vector : ； in, is the learned weight matrix; is the feature dimension of each point; To calculate the attention weight, first calculate and The similarity score between : ； in, is the result of scaling the dot product; The function is used to normalize the score; T is the transpose symbol; Finally, the attention weight Applied to a vector of values , and get the final output : 。 4. The method for determining geometric features of bridge components based on a hybrid model according to claim 2, characterized in that: The training process of the bridge component recognition model is specifically as follows: (1) Process the bridge component point cloud data in batches, where each batch contains multiple points, each point contains the coordinate position (x, y, z) and additional feature values ​​for each point; The additional eigenvalues ​​include normal vector, surface direction, curvature, and concavity; (2) Encoding the batch-processed bridge component point cloud data; (3) Standardize the encoded bridge component point cloud data; (4) Enhance the point cloud data of bridge components, 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 enhanced 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 consists of 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 the nearest neighbor points, construct local areas, each local area contains a central point and its neighbor points; c. Feature extraction: Use a PointNet module to extract features from each local area; (6) The global feature vector is then processed through an MLP layer to output the class probability distribution of the object; (7) Use the cross entropy loss function to train the model by minimizing the difference between the predicted category and the true category.

5. The method for determining geometric features of bridge components based on a hybrid model according to claim 1, characterized in that: The segmentation process of the bridge component is specifically as follows: randomly selecting a plurality of seed points as the starting points of the segmentation in the bridge component point cloud data after the bridge component recognition model recognizes the bridge component; Starting from the initial seed point, the adjacent unlabeled points are checked step by step to see if they satisfy the consistency similarity criterion of the normal line; If the criteria are met, the point is added to the current area and continues to expand until there are no points that meet the criteria to be added, and the initial segmentation area is obtained; When the expansion of an area is completed, a new seed point is selected to continue segmentation until the preset segmentation target is reached; finally, the point cloud is divided into several non-overlapping areas or categories, and the category to which each point belongs is output to complete the segmentation of bridge components.

6. The method for determining geometric features of bridge components based on a hybrid model according to claim 1, characterized in that: The construction process of the dual-mode fusion representation model of point cloud and wireframe is specifically as follows: S201, extracting geometric boundary information of the segmented component parts; S202, generating a skeleton based on the extracted boundary information, wherein the skeleton includes a geometric center and a skeleton line of the component; S203, performing detailed modeling of key parts in the assembly process based on the generated skeleton information and the bridge component point cloud data; S204, smoothing and simplifying the refined modeling results; obtaining a dual-mode fusion representation model including point cloud and wireframe.

7. A method for determining geometric features of bridge components based on a hybrid model according to claim 6, characterized in that: The refined modeling process described in S203 is specifically as follows: based on the generated skeleton information combined with the point cloud data of the bridge component, the coordinates of the key points are marked at the branch points and end points of the component, and their normal vectors and curvatures are extracted; then, the skeleton information is aligned with the point cloud data to ensure that the two are in the same coordinate system, and then the local point cloud data of the key points is extracted; the local point cloud data is preprocessed to remove noise and outliers; the local point cloud is fitted with a surface to generate a refined geometric model, and a continuous surface is generated based on the point cloud normal vector; detailed information is added to the key parts, and boundary points are extracted from the point cloud to enhance the geometric performance of the joints; 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, the local point cloud is re-extracted and the above steps are repeated; finally, the specific parameters of the refined geometric model are output.

8. The method for determining geometric features of bridge components based on a hybrid model according to claim 1, characterized in that: The specific process of S3 is as follows: S301, mapping the three-dimensional point cloud data of the dual-mode fusion representation model including the point cloud and the wireframe to two dimensions; specifically, first constructing a graph to represent the local relationship in the three-dimensional point cloud data of the dual-mode fusion representation model including the point cloud and the wireframe, and then obtaining a low-dimensional representation by calculating the eigenvector of the Laplacian matrix of the graph; S302, extracting edge information of bridge components; S303. Determine the size of bridge components.

9. A method for determining geometric features of bridge components based on a hybrid model according to claim 8, characterized in that: The specific process of S302 is: extracting the edge lines of the component based on the Alpha shape algorithm, and flexibly adjusting the extraction accuracy by setting different Alpha values; smoothing the extracted edge data, and using the adjacent point connection algorithm to repair the missing edge connections to ensure the integrity and continuity of the edge model, and finally obtaining the optimized edge model.

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

Images