Three-dimensional structure reconstruction system based on artificial intelligence topological optimization

Abstract

The invention relates to the technical field of three-dimensional reconstruction, and discloses a three-dimensional structure reconstruction system based on artificial intelligence topological optimization, and the system is characterized in that a data extraction module is used for carrying out region screening according to a coordinate point distribution density value and junction surface grid boundary data, and generating a component partition determination value; the structure identification module is used for comparing, judging whether conditions are met or not, and obtaining boundary regularity coefficients; the topology adjustment module is used for screening a coordinate set of which the deviation value is within a relative boundary range and the height variation is smaller than a critical height variation coefficient to obtain a boundary topology adjustment coefficient; the feature reconstruction module is used for calculating the component form deviation degree according to the product of the node distribution density and the angle change value and generating a reconstructed feature prediction value; the reconstruction prediction module is used for calculating an equilibrium factor between a component structure state value and a credible level coefficient, and obtaining a structure reconstruction trend degree; according to the invention, the need of manual intervention is reduced, and the cost is saved.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction technology, and more specifically to a three-dimensional structure reconstruction system based on artificial intelligence topology optimization. Background Technology

[0002] 3D reconstruction technology aims to collect information from various sensors and data sources to construct a 3D digital model of the physical world. Currently, commonly used core technologies include computer vision, laser scanning, and stereo photogrammetry. While these technologies can capture spatial information and appearance details of complex environments and objects, and generate 3D models for virtual reality, Building Information Modeling (BIM), industrial design, and other fields, they face numerous challenges in practical applications. Traditional 3D reconstruction technologies demand significant time and computing resources when processing large-scale or particularly complex datasets. For example, in the process of building a Building Information Modeling (BIM) system, traditional methods are slow in processing massive amounts of building data, leading to project delays. In industrial design, the high consumption of computing resources increases design costs when processing complex component data. Moreover, traditional methods struggle to adapt to rapidly changing task requirements and lack accuracy, easily resulting in inaccurate designs. Furthermore, traditional 3D reconstruction methods have low levels of automation and intelligence; each task requires manual adjustments and corrections by professionals, which is complex and error-prone. This severely limits the widespread application of 3D reconstruction technology in modern application scenarios that require rapid and continuous data processing. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a three-dimensional structure reconstruction system based on artificial intelligence topology optimization.

[0004] This invention provides a three-dimensional structure reconstruction system based on artificial intelligence topology optimization, the system comprising:

[0005] The data extraction module is used to acquire point cloud image data, surface mesh boundary data and spatial scale parameters of building components. Based on the coordinate point distribution density values ​​in the component point cloud image data, combined with the surface mesh boundary data, the module performs region filtering, selects component regions with closed boundaries and data continuity as target regions, and generates component partitioning judgment values.

[0006] The structure recognition module is used to call the boundary line data of each sub-region in the surface mesh boundary data according to the component partition judgment value, compare the position parameters, corresponding edge length and number of corners of the gap area in the boundary line, determine whether the boundary continuity and regularity judgment conditions are met, and obtain the boundary regularity coefficient.

[0007] The topology adjustment module is used to call spatial coordinate points in point cloud image data according to the boundary regularity coefficient, calculate the relative offset distance of each coordinate point by the height difference and the boundary contour, and filter the set of coordinates whose offset value is within the relative boundary range and whose height change is less than the critical height change coefficient to obtain the boundary topology adjustment coefficient.

[0008] The feature reconstruction module is used to call the boundary contour lines of each component based on the boundary topology adjustment coefficient, obtain the number of nodes, node distribution density and the corresponding angle change value of each node, calculate the degree of deviation of the component shape by the product of node distribution density and angle change value, and generate the reconstructed feature prediction value.

[0009] The reconstruction prediction module is used to call the reconstruction feature prediction value, calculate the difference between the prediction value and the standard feature value of each component category, calculate the balance factor between the component structural state value and the confidence level coefficient based on the reconstruction confidence level coefficient corresponding to the interval of the difference value, and establish the overall structural reconstruction trend degree of the component by averaging the balance factor values ​​of each component.

[0010] Optionally, in a first implementation of the present invention, the data extraction module includes:

[0011] The traversal submodule is used to set the filtering criteria and traverse every coordinate point in the point cloud image data of building components to obtain the initially filtered areas.

[0012] The determination submodule is used to determine the associated boundary line for each dense point in the surface mesh boundary data within the initially selected area;

[0013] The first calculation submodule is used to record each closed boundary region that may meet the requirements. It randomly selects two adjacent points within the region and calculates the distance between the two points. If the distance is less than or equal to the continuity distance standard, it continues to check the distances of other adjacent points within the region.

[0014] The determination submodule is used to determine whether the data in a region is continuous when the distance between all adjacent points in the region is less than or equal to the continuity distance criterion.

[0015] The generation submodule is used to take the component region with closed boundaries and data continuity as the target region and generate component partitioning judgment values.

[0016] Optionally, in a second implementation of the present invention, the setting of the screening criteria includes:

[0017] Set a standard value for judging the density of coordinate points, and denote it as the density judgment standard; set a minimum value for measuring the size of the closed boundary, and denote it as the minimum boundary standard; set a maximum distance for judging whether the data is continuous, i.e., the continuity distance standard.

[0018] Optionally, in a third implementation of the present invention, the structure recognition module includes:

[0019] The extraction submodule is used to extract the boundary line data of each sub-region from the surface mesh boundary data based on the component partitioning judgment value.

[0020] The first identification submodule is used to process the extracted boundary line data using an edge detection algorithm, identify discontinuous parts in the boundary line, determine them as gap areas, record the specific position of each gap area on the boundary line, and obtain the position parameters of the gap area.

[0021] The measurement submodule is used to measure each segment of the boundary line sequentially, starting from the starting point of the boundary line. After traversing the entire boundary line, the lengths of all segments are accumulated to obtain the total edge length of the corresponding boundary line.

[0022] The move submodule is used to move along the boundary line, form the position of the corner where there is a significant change in direction, and determine the number of corners along the boundary line;

[0023] The sorting submodule is used to sort the obtained gap region location parameters, edge lengths and number of corners into a dataset containing boundary feature information;

[0024] The first comparison submodule is used to compare the organized boundary feature dataset with the pre-defined boundary continuity and regularity determination conditions.

[0025] The input submodule is used to input the boundary feature data after comparison under preset conditions into the support vector machine algorithm, classify and judge the boundary features, and output the evaluation results to obtain the boundary regularity coefficient.

[0026] Optionally, in a fourth implementation of the present invention, the first identification submodule includes:

[0027] The extracted boundary line data are convolved using gradient operators of different scales to generate multi-scale gradient maps.

[0028] Calculate the gradient direction of each pixel and divide each pixel into several directional intervals. Perform gradient magnitude consistency checks on adjacent pixels within the same directional interval. If the difference in gradient direction magnitude between adjacent pixels exceeds a threshold, it is marked as a suspicious gap point.

[0029] By introducing a time dimension and constructing a spatiotemporal correlation matrix, for each suspicious gap point in the current frame, we check whether there are similar gradient features at its corresponding position in adjacent frames. If so, we enhance its confidence as a real gap.

[0030] Dynamic thresholds are calculated based on the gradient statistical characteristics of local regions, and adaptive threshold segmentation is performed on the gradient magnitude image to generate a binary edge map.

[0031] Morphological analysis is performed on the fracture edges in the binary edge map. The edge centerline is extracted through skeletonization operation, the curvature change rate of the centerline is calculated, and suspicious gap regions are searched near the curvature abrupt change points. Geometric features are verified for each suspicious gap region.

[0032] For a confirmed gap region, topological prediction is performed using the direction information of the edge segments at both ends. By fitting the potential edge paths of the gap region with Bézier curves, the complete boundary topology is reconstructed to identify the gap region.

[0033] Optionally, in a fifth implementation of the present invention, the topology adjustment module includes:

[0034] The hierarchical indexing submodule is used to perform hierarchical indexing of 3D point cloud data based on boundary regularity coefficients, adding multi-scale geometric feature descriptors to each spatial coordinate point to form an enhanced point cloud feature space.

[0035] The adjustment submodule is used to generate an adaptive reference plane based on the geometric center and main direction of the boundary contour line, and dynamically adjust the tolerance of the reference plane according to the boundary regularity coefficient.

[0036] The second identification submodule is used to construct a spatiotemporal cube to perform contextual analysis on each coordinate point, extract three-dimensional spatial features and time series features, and automatically identify the feature dimensions of boundary topology adjustments through an attention mechanism.

[0037] The fusion submodule is used to perform dual-path CNN feature fusion, generate joint feature representation, derive critical height change coefficient and relative boundary range parameter based on the statistical characteristics of the input 3D point cloud data, construct a decision network to classify each point, and output probability distribution;

[0038] The filtering submodule is used to establish a point cloud structure, where vertices are coordinate points and edges are the topological relationships between points. It filters the set of coordinates in the point cloud structure whose offset values ​​are within the relative boundary range and whose height change is less than the critical height change coefficient, thus obtaining the boundary topology adjustment coefficient.

[0039] Optionally, in the sixth implementation of the present invention, the dual-path CNN feature fusion includes:

[0040] Extract the structural information of the 3D point cloud data in the vertical direction to obtain the height difference path. The structural information includes at least the change features, capturing steps and cracks.

[0041] Analyze the horizontal offset relationship between 3D point cloud data and boundary contours, identify blurred edge areas, and obtain the offset path;

[0042] The feature fusion module of the CNN network nonlinearly combines the height difference path and the offset path to generate a joint feature representation.

[0043] Optionally, in a seventh implementation of the present invention, the feature reconstruction module includes:

[0044] The sorting submodule is used to sort the importance of component boundary contours according to the boundary topology adjustment coefficient and establish a spatial index structure for the contours.

[0045] A submodule is formed to uniformly sample the initial node set on the contour line, record the node coordinates and normal directions, and apply an adaptive sampling strategy to increase the sampling density in areas with large curvature changes to form a multi-scale node representation.

[0046] The statistics submodule is used to define a local neighborhood centered on each node, count the number of nodes in the neighborhood, calculate the normalized node distribution density value based on the neighborhood radius and the number of nodes, and apply Gaussian smoothing to eliminate local density fluctuation noise.

[0047] The filtering submodule is used to calculate the angle of change of the tangent direction between adjacent nodes, which is used as the node angle change value. A two-way angle calculation strategy is used to perform gradient filtering on the angle change value.

[0048] The weighted submodule is used to multiply the node distribution density with the angle change value point by point. It applies a spatial attention mechanism to weight the product results of key areas and aggregates the product values ​​of all nodes through regional integration to obtain the overall shape deviation.

[0049] The mapping submodule is used to map the morphological deviation to a preset discrete level, construct a multi-dimensional feature vector by combining the geometric features of the contour, apply principal component analysis to reduce dimensionality, and generate reconstructed feature prediction values.

[0050] Optionally, in the eighth implementation of the present invention, if there are multiple frames of time-series data, a correspondence between the current frame node and the historical frame is established, the time derivative of the node angle change is calculated, a spatiotemporal feature vector is constructed, and Kalman filtering is applied to predict the future position of the node.

[0051] Optionally, in a ninth implementation of the present invention, the reconstruction prediction module includes:

[0052] The second comparison submodule is used to classify and organize the reconstructed feature prediction values ​​according to the component category. For each component, its reconstructed feature prediction value is compared with the corresponding component category morphological standard feature value in turn, and the differences of all feature dimensions are summarized to obtain the comprehensive difference value of the component.

[0053] The second calculation submodule is used to set an initial component structural state value for each component based on the reconstruction confidence level coefficient corresponding to the interval where the difference value is located, and to calculate the component structural state value and reconstruction confidence level coefficient of each component to obtain the balance factor between the component structural state value and the confidence level coefficient.

[0054] The summation submodule is used to sum the equilibrium factor values ​​of all components, divide by the total number of components, and obtain the average value to determine the overall structural reconstruction trend of the components.

[0055] In the technical solution provided by this invention, point cloud image data, surface mesh boundary data, and spatial scale parameters of building components are acquired. Based on the coordinate point distribution density values ​​in the component point cloud image data, combined with the surface mesh boundary data, region filtering is performed. Component regions with closed boundaries and data continuity are selected as target regions, generating component partitioning judgment values. Based on the component partitioning judgment values, the boundary line data of each sub-region in the surface mesh boundary data is called. The position parameters, corresponding edge lengths, and number of corners of the gap areas in the boundary lines are compared to determine whether the boundary continuity and regularity judgment conditions are met, obtaining the boundary regularity coefficient. Based on the boundary regularity coefficient, spatial coordinate points in the point cloud image data are called. Each coordinate point is calculated simultaneously based on its height difference and the relative offset distance of the boundary contour. The set of coordinates whose offset values ​​are within the relative boundary range and whose height change is less than the critical height change coefficient is selected to obtain the boundary topology adjustment coefficient. Based on the boundary topology adjustment coefficient, various... The system obtains the number of nodes, node distribution density, and corresponding angle change values ​​of each node along the component boundary contour line. It then calculates the component morphological deviation by multiplying the node distribution density by the angle change value, generating a predicted reconstruction feature value. This predicted value is then compared with the standard morphological feature values ​​for each component category. Based on the reconstruction confidence level coefficient corresponding to the interval of the difference value, an equilibrium factor is calculated between the component structural state value and the confidence level coefficient. The equilibrium factor values ​​of each component are averaged to establish the overall structural reconstruction trend degree of the component, thus obtaining the structural reconstruction trend degree. This invention reduces the need for manual intervention by automatically adjusting and optimizing data processing steps. When reconstructing the same type of component multiple times, the system continuously optimizes the reconstruction results to make them closer to the actual physical structure, greatly enhancing the model's practicality and reliability. When processing 3D reconstruction data of large urban buildings, the required number of computing devices and computation time are significantly reduced, saving costs. Attached Figure Description

[0056] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0057] Figure 1 A schematic diagram of the structure of a three-dimensional structure reconstruction system based on artificial intelligence topology optimization provided in an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the structure of the data extraction module provided in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the structure recognition module provided in an embodiment of the present invention. Detailed Implementation

[0060] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0061] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A schematic diagram of a three-dimensional structure reconstruction system based on artificial intelligence topology optimization provided in this embodiment of the invention. The system includes:

[0062] The data extraction module is used to acquire point cloud image data, surface mesh boundary data and spatial scale parameters of building components. Based on the coordinate point distribution density values ​​in the component point cloud image data, combined with the surface mesh boundary data, the module performs region filtering, selects component regions with closed boundaries and data continuity as target regions, and generates component partitioning judgment values.

[0063] The structure recognition module is used to call the boundary line data of each sub-region in the surface mesh boundary data according to the component partition judgment value, compare the position parameters, corresponding edge length and number of corners of the gap area in the boundary line, determine whether the boundary continuity and regularity judgment conditions are met, and obtain the boundary regularity coefficient.

[0064] The topology adjustment module is used to call spatial coordinate points in point cloud image data according to the boundary regularity coefficient, calculate the relative offset distance of each coordinate point by the height difference and the boundary contour, and filter the set of coordinates whose offset value is within the relative boundary range and whose height change is less than the critical height change coefficient to obtain the boundary topology adjustment coefficient.

[0065] The feature reconstruction module is used to call the boundary contour lines of each component based on the boundary topology adjustment coefficient, obtain the number of nodes, node distribution density and the corresponding angle change value of each node, calculate the degree of deviation of the component shape by the product of node distribution density and angle change value, and generate the reconstructed feature prediction value.

[0066] The reconstruction prediction module is used to call the reconstruction feature prediction value, calculate the difference between the prediction value and the standard feature value of each component category, calculate the balance factor between the component structural state value and the confidence level coefficient based on the reconstruction confidence level coefficient corresponding to the interval of the difference value, and establish the overall structural reconstruction trend degree of the component by averaging the balance factor values ​​of each component.

[0067] In this embodiment, please refer to Figure 2 The data extraction module includes:

[0068] The traversal submodule is used to set the filtering criteria and traverse every coordinate point in the point cloud image data of building components to obtain the initially filtered areas.

[0069] The determination submodule is used to determine the associated boundary line for each dense point in the surface mesh boundary data within the initially selected area;

[0070] The first calculation submodule is used to record each closed boundary region that may meet the requirements. It randomly selects two adjacent points within the region and calculates the distance between the two points. If the distance is less than or equal to the continuity distance standard, it continues to check the distances of other adjacent points within the region.

[0071] The determination submodule is used to determine whether the data in a region is continuous when the distance between all adjacent points in the region is less than or equal to the continuity distance criterion.

[0072] The generation submodule is used to take the component region with closed boundaries and data continuity as the target region and generate component partitioning judgment values.

[0073] In this embodiment, the screening criteria include:

[0074] Set a standard value for judging the density of coordinate points, and denote it as the density judgment standard; set a minimum value for measuring the size of the closed boundary, and denote it as the minimum boundary standard; set a maximum distance for judging whether the data is continuous, i.e., the continuity distance standard.

[0075] In this embodiment, please refer to Figure 3 The structure recognition module includes:

[0076] The extraction submodule is used to extract the boundary line data of each sub-region from the surface mesh boundary data based on the component partitioning judgment value.

[0077] The first identification submodule is used to process the extracted boundary line data using an edge detection algorithm, identify discontinuous parts in the boundary line, determine them as gap areas, record the specific position of each gap area on the boundary line, and obtain the position parameters of the gap area.

[0078] The measurement submodule is used to measure each segment of the boundary line sequentially, starting from the starting point of the boundary line. After traversing the entire boundary line, the lengths of all segments are accumulated to obtain the total edge length of the corresponding boundary line.

[0079] The move submodule is used to move along the boundary line, form the position of the corner where there is a significant change in direction, and determine the number of corners along the boundary line;

[0080] The sorting submodule is used to sort the obtained gap region location parameters, edge lengths and number of corners into a dataset containing boundary feature information;

[0081] The first comparison submodule is used to compare the organized boundary feature dataset with the pre-defined boundary continuity and regularity determination conditions.

[0082] The input submodule is used to input the boundary feature data after comparison under preset conditions into the support vector machine algorithm, classify and judge the boundary features, and output the evaluation results to obtain the boundary regularity coefficient.

[0083] In this embodiment, the first identification submodule includes:

[0084] The extracted boundary line data are convolved using gradient operators of different scales to generate multi-scale gradient maps.

[0085] Calculate the gradient direction of each pixel and divide each pixel into several directional intervals. Perform gradient magnitude consistency checks on adjacent pixels within the same directional interval. If the difference in gradient direction magnitude between adjacent pixels exceeds a threshold, it is marked as a suspicious gap point.

[0086] By introducing a time dimension and constructing a spatiotemporal correlation matrix, for each suspicious gap point in the current frame, we check whether there are similar gradient features at its corresponding position in adjacent frames. If so, we enhance its confidence as a real gap.

[0087] Dynamic thresholds are calculated based on the gradient statistical characteristics of local regions, and adaptive threshold segmentation is performed on the gradient magnitude image to generate a binary edge map.

[0088] Morphological analysis is performed on the fracture edges in the binary edge map. The edge centerline is extracted through skeletonization operation, the curvature change rate of the centerline is calculated, and suspicious gap regions are searched near the curvature abrupt change points. Geometric features are verified for each suspicious gap region.

[0089] For a confirmed gap region, topological prediction is performed using the direction information of the edge segments at both ends. By fitting the potential edge paths of the gap region with Bézier curves, the complete boundary topology is reconstructed to identify the gap region.

[0090] In this embodiment, the topology adjustment module includes:

[0091] The hierarchical indexing submodule is used to perform hierarchical indexing of 3D point cloud data based on boundary regularity coefficients, adding multi-scale geometric feature descriptors to each spatial coordinate point to form an enhanced point cloud feature space.

[0092] The adjustment submodule is used to generate an adaptive reference plane based on the geometric center and main direction of the boundary contour line, and dynamically adjust the tolerance of the reference plane according to the boundary regularity coefficient.

[0093] The second identification submodule is used to construct a spatiotemporal cube to perform contextual analysis on each coordinate point, extract three-dimensional spatial features and time series features, and automatically identify the feature dimensions of boundary topology adjustments through an attention mechanism.

[0094] The fusion submodule is used to perform dual-path CNN feature fusion, generate joint feature representation, derive critical height change coefficient and relative boundary range parameter based on the statistical characteristics of the input 3D point cloud data, construct a decision network to classify each point, and output probability distribution;

[0095] The filtering submodule is used to establish a point cloud structure, where vertices are coordinate points and edges are the topological relationships between points. It filters the set of coordinates in the point cloud structure whose offset values ​​are within the relative boundary range and whose height change is less than the critical height change coefficient, thus obtaining the boundary topology adjustment coefficient.

[0096] In this embodiment, dual-path CNN feature fusion includes:

[0097] Extract the structural information of the 3D point cloud data in the vertical direction to obtain the height difference path. The structural information includes at least the change features, capturing steps and cracks.

[0098] Analyze the horizontal offset relationship between 3D point cloud data and boundary contours, identify blurred edge areas, and obtain the offset path;

[0099] The feature fusion module of the CNN network nonlinearly combines the height difference path and the offset path to generate a joint feature representation.

[0100] In this embodiment, the feature reconstruction module includes:

[0101] The sorting submodule is used to sort the importance of component boundary contours according to the boundary topology adjustment coefficient and establish a spatial index structure for the contours.

[0102] A submodule is formed to uniformly sample the initial node set on the contour line, record the node coordinates and normal directions, and apply an adaptive sampling strategy to increase the sampling density in areas with large curvature changes to form a multi-scale node representation.

[0103] The statistics submodule is used to define a local neighborhood centered on each node, count the number of nodes in the neighborhood, calculate the normalized node distribution density value based on the neighborhood radius and the number of nodes, and apply Gaussian smoothing to eliminate local density fluctuation noise.

[0104] The filtering submodule is used to calculate the angle of change of the tangent direction between adjacent nodes, which is used as the node angle change value. A two-way angle calculation strategy is used to perform gradient filtering on the angle change value.

[0105] The weighted submodule is used to multiply the node distribution density with the angle change value point by point. It applies a spatial attention mechanism to weight the product results of key areas and aggregates the product values ​​of all nodes through regional integration to obtain the overall shape deviation.

[0106] The mapping submodule is used to map the morphological deviation to a preset discrete level, construct a multi-dimensional feature vector by combining the geometric features of the contour, apply principal component analysis to reduce dimensionality, and generate reconstructed feature prediction values.

[0107] In this embodiment, if there are multiple frames of time-series data, the correspondence between the current frame node and the historical frame is established, the time derivative of the node angle change is calculated, a spatiotemporal feature vector is constructed, and Kalman filtering is applied to predict the future position of the node.

[0108] In this embodiment, the reconstruction prediction module includes:

[0109] The second comparison submodule is used to classify and organize the reconstructed feature prediction values ​​according to the component category. For each component, its reconstructed feature prediction value is compared with the corresponding component category morphological standard feature value in turn, and the differences of all feature dimensions are summarized to obtain the comprehensive difference value of the component.

[0110] The second calculation submodule is used to set an initial component structural state value for each component based on the reconstruction confidence level coefficient corresponding to the interval where the difference value is located, and to calculate the component structural state value and reconstruction confidence level coefficient of each component to obtain the balance factor between the component structural state value and the confidence level coefficient.

[0111] The summation submodule is used to sum the equilibrium factor values ​​of all components, divide by the total number of components, and obtain the average value to determine the overall structural reconstruction trend of the components.

[0112] In this embodiment, the boundary feature data after comparison under preset conditions is normalized to ensure that different feature dimensions have the same scale range, avoiding unreasonable impact on the model due to some features being too large or too small. Based on the boundary feature data, feature engineering is used to expand data features; for example, derived features such as the ratio of edge length to average edge length and the difference between the number of corners and a specific standard number of corners are calculated to enrich the expressive power of the data and provide more information for subsequent classification and judgment. The distribution characteristics of the boundary feature data are analyzed, and various kernel functions are designed, such as polynomial kernel functions, radial basis kernel functions, and Gaussian kernel functions. Appropriate kernel functions or combinations of kernel functions are dynamically selected according to the local density and dimensional characteristics of the data. For regions with relatively concentrated data distribution, kernel functions that can better fit local features are selected; for relatively dispersed data, kernel functions that can map in a wider space are selected. Through this adaptive selection, the adaptability of the model to different data distributions is improved. Dynamic weights are assigned to each boundary feature data sample. Considering factors such as the distance between the sample and the classification hyperplane and the degree of feature difference of the sample, weights are assigned to samples that are far from the hyperplane and have relatively small feature differences. Larger samples are assigned higher weights, and vice versa. This allows the model to focus on samples that have a key impact on determining the classification boundary during training and judgment, improving the model's accuracy. Multiple support vector machine (SVM) classifiers are constructed, each trained on a different subset of features. For example, one classifier might focus on edge length-related features, while another might focus on corner number-related features. During classification and judgment, these multiple classifiers work collaboratively. The judgment results from multiple classifiers are combined, and different confidence weights are assigned to each classifier based on their performance during training. A weighted fusion method is then used to obtain a comprehensive judgment result, enhancing the model's robustness and accuracy. After obtaining the initial boundary regularity coefficients, an online learning mechanism is activated if new boundary feature data is added. The improved SVM model is incrementally trained using the new data, updating the model's parameters. Based on the characteristics of the new data and the model's feedback, the previously set weights and kernel function selection strategies are dynamically adjusted, enabling the model to continuously optimize as the data changes, better adapt to different boundary feature situations, and consistently output more accurate boundary regularity coefficients.

[0113] In this embodiment, the height difference path and the offset distance path capture different features of the point cloud related to the boundary in the vertical and horizontal directions, respectively. The joint feature representation integrates these two aspects of features, which can more comprehensively and holistically reflect the relationship between the coordinate points and the boundary. For example, when processing point cloud data of complex building structures, it can reflect both the structural changes at different heights of the building facade (such as the height difference features caused by window sills, waistlines, etc.) and the offset features of the wall edge caused by construction errors or erosion from the design boundary, providing rich information for subsequent accurate judgment of whether the coordinate points meet the boundary topology adjustment requirements. The fused joint feature representation enables the subsequent decision network to more accurately distinguish different types of points. Through this comprehensive feature representation, the decision network can better identify coordinate points whose offset values ​​are within the relative boundary range and whose height change is less than the critical height change coefficient, reducing misjudgments and omissions. For example... When processing point cloud data with noise interference, relying solely on height difference or offset distance features may misclassify some noisy points as valid points. Joint feature representation, however, can utilize multiple information sources to more accurately eliminate noise and filter out the set of coordinate points that truly meet the conditions, thereby improving the accuracy of boundary topology adjustment coefficient calculation. Joint feature representation can also enhance the model's adaptability to different scenarios and data characteristics. Different 3D reconstruction scenarios, such as industrial component reconstruction and terrain reconstruction, have significantly different features and boundary conditions in point cloud data. Joint feature representation can flexibly integrate features from different aspects, enabling the model to work effectively in various complex situations. For example, in industrial scenarios, minor defects on the surface of equipment may manifest as a combination of height difference and offset features. Joint feature representation can help the model better adapt to such subtle feature changes, accurately perform boundary topology adjustments, and enhance the model's versatility and robustness.

[0114] In this embodiment, contour lines are weighted and sorted according to boundary topology adjustment coefficients. Higher coefficients indicate a more stable topological structure and a greater contribution to the reconstruction results. A multi-index sorting strategy is adopted, considering boundary length, closure, and relevance to neighbors to construct a priority queue for efficient processing of high-priority contour lines. A hybrid index is used for contour lines of varying complexity: quadtree indexes are used for regular regions to improve query efficiency, while KD-tree indexes are used for complex regions to optimize nearest-neighbor search performance. A multi-level index structure is established, with the top-level index quickly locating approximate regions and the bottom-level index accurately matching specific contours. Based on the contour line length and a preset sampling interval, an initial node set is generated. An equal-arc-length sampling method is used to ensure uniform distribution of nodes on the curve, recording the 3D coordinates and local surface normal direction of each node. The curvature values ​​of each point on the contour line are calculated, identifying regions with drastic curvature changes (such as corners and breakpoints). Sampling density is increased in high-curvature regions, and a binary insertion method is used to refine the node distribution, constructing a multi-resolution node hierarchy. Coarse-scale nodes capture the overall morphology, while fine-scale nodes preserve detailed features. A variable-radius neighborhood is defined centered on each node, with the radius inversely proportional to the local curvature. The number of nodes within the neighborhood is counted, and the original density value is calculated based on the neighborhood volume. A kernel density estimation method is applied to smooth the density distribution. The density value is normalized using the z-score method to eliminate scale effects. Gaussian filtering is applied to eliminate local density fluctuations while preserving the macroscopic density change trend. A density threshold is introduced to filter outouts and low-density noise regions. The tangent direction change angle between adjacent nodes is calculated, considering both clockwise and counterclockwise directions. A sliding window method is used to calculate the angle change rate, with the window size adapting to the local curvature. The angle change value is periodically processed and mapped to the [0,π] interval. The Sobel operator is applied to calculate the gradient of the angle change to highlight abrupt morphological regions. A direction-sensitive filter is designed to preserve angle changes perpendicular to the main contour direction. An angle change pyramid is constructed to capture morphological features at different scales.The normalized node distribution density is multiplied point-by-point by the angle change value to generate an initial deviation map. A spatial attention mechanism is applied, assigning higher weights to key regions (such as corners and endpoints). Topological constraints are introduced to penalize deviations that disrupt the connectivity of the original contour. The contour line is divided into multiple semantic regions (such as walls and doors / windows), and the deviation of each region is calculated. A weighted integral method is used to aggregate the deviations of each region, with weights based on region importance and area, constructing a cumulative deviation distribution map to visually display the spatial distribution of morphological anomalies. Continuous morphological deviation values ​​are mapped to preset discrete levels (such as levels 1-5), using K... The -means clustering method automatically determines the grading threshold, assigns semantic labels (e.g., "normal," "slightly deformed," "severely damaged") to each grade, combines contour geometric features (perimeter, area, roundness) and topological features (number of holes, number of branches) to extract shape descriptors (e.g., Fourier descriptors, Zernike moments) to enhance feature representation, constructs spatiotemporal feature vectors, integrates the changing trends of multi-frame data, calculates the feature covariance matrix, determines the contribution rate of each feature dimension, selects principal components with a cumulative contribution rate exceeding a threshold (e.g., 95%), generates low-dimensional feature vectors, and removes redundant information.

[0115] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional structure reconstruction system based on artificial intelligence topology optimization, characterized in that, The system includes: The data extraction module is used to acquire point cloud image data, surface mesh boundary data and spatial scale parameters of building components. Based on the coordinate point distribution density values ​​in the component point cloud image data, combined with the surface mesh boundary data, the module performs region filtering, selects component regions with closed boundaries and data continuity as target regions, and generates component partitioning judgment values. The structure recognition module is used to call the boundary line data of each sub-region in the surface mesh boundary data according to the component partition judgment value, compare the position parameters, corresponding edge length and number of corners of the gap area in the boundary line, determine whether the boundary continuity and regularity judgment conditions are met, and obtain the boundary regularity coefficient. The topology adjustment module is used to call spatial coordinate points in point cloud image data according to the boundary regularity coefficient, calculate the relative offset distance of each coordinate point by the height difference and the boundary contour, and filter the set of coordinates whose offset value is within the relative boundary range and whose height change is less than the critical height change coefficient to obtain the boundary topology adjustment coefficient. The feature reconstruction module is used to call the boundary contour lines of each component based on the boundary topology adjustment coefficient, obtain the number of nodes, node distribution density and the corresponding angle change value of each node, calculate the degree of deviation of the component shape by the product of node distribution density and angle change value, and generate the reconstructed feature prediction value. The reconstruction prediction module is used to call the reconstruction feature prediction value, calculate the difference between the prediction value and the standard feature value of each component category, calculate the balance factor between the component structural state value and the confidence level coefficient based on the reconstruction confidence level coefficient corresponding to the interval of the difference value, and establish the overall structural reconstruction trend degree of the component by averaging the balance factor values ​​of each component.

2. The three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 1, characterized in that, The data extraction module includes: The traversal submodule is used to set the filtering criteria and traverse every coordinate point in the point cloud image data of building components to obtain the initially filtered areas. The determination submodule is used to determine the associated boundary line for each dense point in the surface mesh boundary data within the initially selected area; The first calculation submodule is used to record each closed boundary region that may meet the requirements. It randomly selects two adjacent points within the region and calculates the distance between the two points. If the distance is less than or equal to the continuity distance standard, it continues to check the distances of other adjacent points within the region. The determination submodule is used to determine whether the data in a region is continuous when the distance between all adjacent points in the region is less than or equal to the continuity distance criterion. The generation submodule is used to take the component region with closed boundaries and data continuity as the target region and generate component partitioning judgment values.

3. The three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 2, characterized in that, The set screening criteria include: Set a standard value for judging the density of coordinate points, and denote it as the density judgment standard; set a minimum value for measuring the size of the closed boundary, and denote it as the minimum boundary standard; set a maximum distance for judging whether the data is continuous, i.e., the continuity distance standard.

4. The three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 1, characterized in that, The structure recognition module includes: The extraction submodule is used to extract the boundary line data of each sub-region from the surface mesh boundary data based on the component partitioning judgment value. The first identification submodule is used to process the extracted boundary line data using an edge detection algorithm, identify discontinuous parts in the boundary line, determine them as gap areas, record the specific position of each gap area on the boundary line, and obtain the position parameters of the gap area. The measurement submodule is used to measure each segment of the boundary line sequentially, starting from the starting point of the boundary line. After traversing the entire boundary line, the lengths of all segments are accumulated to obtain the total edge length of the corresponding boundary line. The move submodule is used to move along the boundary line, form the position of the corner where there is a significant change in direction, and determine the number of corners along the boundary line; The sorting submodule is used to sort the obtained gap region location parameters, edge lengths and number of corners into a dataset containing boundary feature information; The first comparison submodule is used to compare the organized boundary feature dataset with the pre-defined boundary continuity and regularity determination conditions. The input submodule is used to input the boundary feature data after comparison under preset conditions into the support vector machine algorithm, classify and judge the boundary features, and output the evaluation results to obtain the boundary regularity coefficient.

5. The three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 4, characterized in that, The first identification submodule includes: The extracted boundary line data are convolved using gradient operators of different scales to generate multi-scale gradient maps. Calculate the gradient direction of each pixel and divide each pixel into several directional intervals. Perform gradient magnitude consistency checks on adjacent pixels within the same directional interval. If the difference in gradient direction magnitude between adjacent pixels exceeds a threshold, it is marked as a suspicious gap point. By introducing a time dimension and constructing a spatiotemporal correlation matrix, for each suspicious gap point in the current frame, we check whether there are similar gradient features at its corresponding position in adjacent frames. If so, we enhance its confidence as a real gap. Dynamic thresholds are calculated based on the gradient statistical characteristics of local regions, and adaptive threshold segmentation is performed on the gradient magnitude image to generate a binary edge map. Morphological analysis is performed on the fracture edges in the binary edge map. The edge centerline is extracted through skeletonization operation, the curvature change rate of the centerline is calculated, and suspicious gap regions are searched near the curvature abrupt change points. Geometric features are verified for each suspicious gap region. For a confirmed gap region, topological prediction is performed using the direction information of the edge segments at both ends. By fitting the potential edge paths of the gap region with Bézier curves, the complete boundary topology is reconstructed to identify the gap region.

6. The three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 1, characterized in that, The topology adjustment module includes: The hierarchical indexing submodule is used to perform hierarchical indexing of 3D point cloud data based on boundary regularity coefficients, adding multi-scale geometric feature descriptors to each spatial coordinate point to form an enhanced point cloud feature space. The adjustment submodule is used to generate an adaptive reference plane based on the geometric center and main direction of the boundary contour line, and dynamically adjust the tolerance of the reference plane according to the boundary regularity coefficient. The second identification submodule is used to construct a spatiotemporal cube to perform contextual analysis on each coordinate point, extract three-dimensional spatial features and time series features, and automatically identify the feature dimensions of boundary topology adjustments through an attention mechanism. The fusion submodule is used to perform dual-path CNN feature fusion, generate joint feature representation, derive critical height change coefficient and relative boundary range parameter based on the statistical characteristics of the input 3D point cloud data, construct a decision network to classify each point, and output probability distribution; The filtering submodule is used to establish a point cloud structure, where vertices are coordinate points and edges are the topological relationships between points. It filters the set of coordinates in the point cloud structure whose offset values ​​are within the relative boundary range and whose height change is less than the critical height change coefficient, thus obtaining the boundary topology adjustment coefficient.

7. The three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 6, characterized in that, The dual-path CNN feature fusion includes: Extract the structural information of the 3D point cloud data in the vertical direction to obtain the height difference path. The structural information includes at least the change features, capturing steps and cracks. Analyze the horizontal offset relationship between 3D point cloud data and boundary contours, identify blurred edge areas, and obtain the offset path; The feature fusion module of the CNN network nonlinearly combines the height difference path and the offset path to generate a joint feature representation.

8. The three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 1, characterized in that, The feature reconstruction module includes: The sorting submodule is used to sort the importance of component boundary contours according to the boundary topology adjustment coefficient and establish a spatial index structure for the contours. A submodule is formed to uniformly sample the initial node set on the contour line, record the node coordinates and normal directions, and apply an adaptive sampling strategy to increase the sampling density in areas with large curvature changes to form a multi-scale node representation. The statistics submodule is used to define a local neighborhood centered on each node, count the number of nodes in the neighborhood, calculate the normalized node distribution density value based on the neighborhood radius and the number of nodes, and apply Gaussian smoothing to eliminate local density fluctuation noise. The filtering submodule is used to calculate the angle of change of the tangent direction between adjacent nodes, which is used as the node angle change value. A two-way angle calculation strategy is used to perform gradient filtering on the angle change value. The weighted submodule is used to multiply the node distribution density with the angle change value point by point. It applies a spatial attention mechanism to weight the product results of key areas and aggregates the product values ​​of all nodes through regional integration to obtain the overall shape deviation. The mapping submodule is used to map the morphological deviation to a preset discrete level, construct a multi-dimensional feature vector by combining the geometric features of the contour, apply principal component analysis to reduce dimensionality, and generate reconstructed feature prediction values.

9. A three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 8, characterized in that, If multiple frames of time-series data exist, establish the correspondence between the current frame node and the historical frames, calculate the time derivative of the node angle change, construct the spatiotemporal feature vector, and apply Kalman filtering to predict the future position of the node.

10. A three-dimensional structure reconstruction system based on artificial intelligence topology optimization as described in claim 1, characterized in that, The reconstruction prediction module includes: The second comparison submodule is used to classify and organize the reconstructed feature prediction values ​​according to the component category. For each component, its reconstructed feature prediction value is compared with the corresponding component category morphological standard feature value in turn, and the differences of all feature dimensions are summarized to obtain the comprehensive difference value of the component. The second calculation submodule is used to set an initial component structural state value for each component based on the reconstruction confidence level coefficient corresponding to the interval where the difference value is located, and to calculate the component structural state value and reconstruction confidence level coefficient of each component to obtain the balance factor between the component structural state value and the confidence level coefficient. The summation submodule is used to sum the equilibrium factor values ​​of all components, divide by the total number of components, and obtain the average value to determine the overall structural reconstruction trend of the components.

Images