A method and system for constructing a three-dimensional model of an overpass

Abstract

The application provides a method and system for constructing a three-dimensional model of an overpass, and relates to the technical field of three-dimensional model construction. The application obtains structural unit point clouds by performing semantic segmentation and instance segmentation on original point cloud data, and constructs an overpass structure topology graph in accordance with the structural unit point clouds. The standard reference geometric information of each node is determined based on the topology graph, missing areas are identified, and the structure unit nodes that are occluded and the occlusion types are determined. Incomplete point clouds are input into a pre-trained conditional generative point cloud completion model for completion to generate complete point clouds. Finally, the parameterized three-dimensional models of each node are generated by matching and fitting with parameterized structure templates, and the complete overpass three-dimensional semantic model is assembled in accordance with the topological relationship. The application combines topological reasoning and conditional generative completion, effectively restores the complete geometric shape of the occluded area, and greatly improves the occlusion repair accuracy and the model completeness.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional model construction technology, specifically to a method and system for constructing a three-dimensional model of an overpass. Background Technology

[0002] With the increasing complexity of urban transportation infrastructure, large-scale interchanges have become a key and challenging area for urban 3D modeling. Airborne LiDAR technology can quickly acquire high-precision 3D point cloud data of large-scale interchanges, providing an important data foundation for the 3D reconstruction of interchanges.

[0003] In existing technologies, the method for reconstructing a 3D model of a large and complex overpass using airborne LiDAR data, disclosed in CN105787921A, employs a segmentation-then-modeling strategy. First, it extracts the overpass point cloud from the airborne LiDAR data using inverse iterative mathematical morphological filtering and area information. Then, it segments the overpass point cloud based on bridge surface connectivity. Next, it uses a centerline perpendicular scanning method to segment and extract connected bridge surfaces without bifurcations or intersections and with consistent width as structural units. Based on this, it detects and repairs occlusion structures. Finally, it combines bridge surface width information to complete the 3D model reconstruction. This type of method has achieved initial success in automating the 3D reconstruction of overpasses, enabling the division of complex intersections and bifurcations into multiple simpler segments for divide-and-conquer processing.

[0004] However, the existing technologies still have the following significant shortcomings: Firstly, in the point cloud segmentation stage, existing methods mainly rely on manually designed rules such as connectivity and centerline perpendicularity scanning for structural unit extraction. This results in poor adaptability to point cloud density variations and noise, and the segmentation accuracy is highly dependent on manually set parameters and thresholds, making end-to-end automated segmentation difficult. Secondly, in the occlusion detection and repair stage, existing methods only create circular buffer zones at the endpoints of the structural unit's centerline. Occlusion is detected by judging the elevation information of the point cloud within the buffer zone. This local detection method can only detect occlusion near the endpoints of the structural unit, and cannot effectively identify situations where the middle area of ​​the structural unit is completely covered by the upper layer, let alone accurately determine the range and type of occlusion. Thirdly, in the repair stage, existing methods mainly rely on the one-dimensional continuity of the road centerline for interpolation repair. This can only restore the direction of the centerline but cannot restore the complete two-dimensional surface morphology of the occluded area (such as bridge width, cross slope, and longitudinal slope), resulting in limited accuracy of the repaired model. Furthermore, existing methods primarily generate triangular mesh geometric models, lacking semantic information such as semantic category labels for structural units and topological connections, making them difficult to directly apply to engineering scenarios requiring semantic 3D models, such as BIM, high-precision maps, and traffic simulation.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for constructing a three-dimensional model of an overpass, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing a 3D model of an overpass, comprising the following steps: Semantic segmentation and instance segmentation are performed on the original point cloud data to obtain multiple structural unit point clouds with semantic category labels and instance labels; Based on the instance labels, each structural unit point cloud is treated as a structural unit node. The topological connection relationship between structural unit nodes is established according to the spatial location of each structural unit node, and the topological map of the overpass structure is constructed. Based on the topology diagram of the overpass structure, the standard reference geometric information corresponding to each structural unit node is determined. The point cloud of each structural unit is compared with the corresponding standard reference geometric information to identify the regions in which the point cloud of each structural unit is missing relative to the corresponding standard reference geometric information. Based on the spatial position of the region in the corresponding structural unit point cloud and the topological connection relationship of the corresponding structural unit node in the overpass structure topology diagram, the occluded structural unit node and its occlusion type are determined. Based on the occlusion type, the incomplete structural unit point cloud of the occluded structural unit node is taken as input, and the spatial position and geometric shape of each structural unit node adjacent to the occluded structural unit node in the overpass structural topology map are used as constraints. The inference is performed in the pre-trained conditional generative point cloud completion model to generate the complete structural unit point cloud of the occluded structural unit node. The complete structural unit point cloud is matched with templates in the preset parametric structural template library. The template with the highest matching degree is selected and the corresponding geometric control parameters are fitted to generate parametric 3D models of each structural unit node. Then, according to the topological connection relationship, each parametric 3D model is assembled according to its corresponding spatial position to form a complete 3D semantic model of the overpass.

[0008] Furthermore, the original point cloud data is airborne lidar point cloud data; the semantic category labels include bridge deck category, ramp category, pier category, and guardrail category; The semantic segmentation and instance segmentation are achieved through a pre-trained 3D point cloud segmentation network. The 3D point cloud segmentation network extracts and classifies the semantic features of the original point cloud data point by point, and performs instance clustering based on the classification results and the spatial connectivity between point clouds, so as to distinguish the geometric regions of point clouds that are not connected under the same semantic category into different instances. The semantic segmentation and instance segmentation specifically include: first, performing point-by-point semantic classification on the original point cloud data and assigning a semantic category label to each point cloud data point; then, for the point cloud data that has been assigned semantic category labels, performing density-based connectivity clustering based on the spatial location of each point cloud data point, dividing the point cloud set that belongs to the same semantic category and is spatially continuous into an independent structural unit instance and assigning a corresponding instance label.

[0009] Furthermore, the topological connection relationship includes connected edges representing structural intersections and forks, and intersecting edges representing overlapping upper and lower spatial layers; the structural intersection refers to the topological connection relationship in which multiple structural unit nodes converge to one structural unit node along the driving direction, and the fork refers to the topological connection relationship in which one structural unit node branches into multiple structural unit nodes along the driving direction. The method for establishing the connecting edge is as follows: when the projections of two structural unit nodes in space overlap, and the structural unit point clouds corresponding to the two structural unit nodes in the overlapping area are connected or intersecting in spatial position, a connecting edge is established between the two structural unit nodes. The method for establishing the cross edge is as follows: when the projections of two structural unit nodes in space overlap, and the point clouds of the structural unit nodes corresponding to the two structural unit nodes in the overlapping area have vertical spacing and do not intersect each other, a cross edge is established between the two structural unit nodes. Each structural unit node is associated with and stores the following attribute information: the spatial bounding box of the structural unit point cloud corresponding to the structural unit node, the structural orientation description information, the elevation description information, and the corresponding semantic category label; When the semantic category label is a bridge deck category or a ramp category, the structural orientation description information includes the centerline orientation or the trajectory line orientation; When the semantic category label is the bridge pier category, the structural orientation description information includes the vertical axis direction; When the semantic category label is the guardrail category, the structural orientation description information includes the horizontal trajectory orientation; The elevation description information includes elevation range or centerline elevation profile data.

[0010] Furthermore, the standard reference geometry information is determined based on the semantic category labels of the structural unit nodes: When the semantic category label of a structural unit node is a bridge deck category or a ramp category, the standard reference geometric information includes the reference centerline orientation, reference width variation range, and reference elevation variation range corresponding to the structural unit. When the semantic category label is a bridge pier, the standard reference geometric information includes the reference vertical axial orientation and reference cross-sectional dimension range corresponding to the structural unit; When the semantic category label is the guardrail category, the standard reference geometry information includes the reference horizontal trajectory and reference height profile range corresponding to the structural unit; For each point cloud data point in each structural unit point cloud, calculate the degree of spatial deviation between its spatial position and the standard position corresponding to the standard reference geometric information; identify the continuous region formed by point cloud data points whose spatial deviation exceeds the preset tolerance range as a missing region.

[0011] Furthermore, the occlusion types include: When the spatial location of the missing region in the corresponding structural unit point cloud completely overlaps with the spatial projection area of ​​any structural unit node connected by a cross edge in the overpass structural topology map, it is determined to be a full occlusion type. When the missing region and the spatial projection area of ​​at least one structural unit node connected by intersecting edges in the overpass structure topology map only partially overlap, and there is no complete overlap with the spatial projection area of ​​any structural unit node, it is determined to be a partial occlusion type; the spatial projection area is the vertical projection coverage of the structural unit point cloud corresponding to each structural unit node on the horizontal plane, and its determination is based on the projection range of the spatial bounding box stored by each structural unit node on the horizontal plane. When the continuity of the connected edges of an obscured structural unit node in the overpass structural topology diagram is interrupted by a missing region, it is determined to be a structural break obscuration type. The determination method for the interruption of the continuity of the connected edges by a missing region is as follows: when the missing region divides the corresponding structural unit point cloud into at least two spatially unconnected point cloud segments along the driving direction, and the upstream and downstream structural unit nodes connected by the obscured structural unit node in the overpass structural topology diagram through connected edges are located exactly within these two point cloud segments, it is determined that the continuity of the connected edges is interrupted by a missing region. For full occlusion, the missing region spans the entire width of the corresponding structural unit point cloud; for partial occlusion, the missing region covers only a portion of the width or length of the corresponding structural unit point cloud; for structural break occlusion, the occluded structural unit node is connected to the upstream and downstream structural unit nodes respectively in the overpass structural topology diagram through connecting edges; the missing region divides the structural unit point cloud corresponding to the occluded structural unit node into at least two spatially unconnected point cloud segments.

[0012] Furthermore, the spatial location includes the spatial coordinates of each adjacent structural unit node, and the geometric shape includes the structural unit width, structural unit orientation vector, and corresponding semantic category label of each adjacent structural unit node; The conditional generative point cloud completion model is a deep learning network model built on a conditional generative adversarial network architecture or a conditional diffusion model architecture. It employs differentiated completion strategies based on different occlusion types, specifically including: When the occlusion type is full occlusion, the conditional generative point cloud completion model uses the spatial position and geometric shape of adjacent structural unit nodes as the only morphological reference to generate the complete structural unit point cloud of the occluded structural unit node. When the occlusion type is partial occlusion, the conditional generative point cloud completion model uses the geometry of the missing part in the incomplete structural unit point cloud of the occluded structural unit node as a reference, and combines the spatial position and geometry of the adjacent structural unit node to perform local completion and generate a complete structural unit point cloud. When the occlusion type is structural fracture occlusion, the conditional generative point cloud completion model takes the point cloud segments on both sides of the fracture of the occluded structural unit node as the reference, and the spatial position and geometry of the adjacent structural unit node as the guide to generate the intermediate transition segment point cloud connecting the two side point cloud segments.

[0013] Furthermore, the pre-training method for the conditional generative point cloud completion model is as follows: Multiple complete point cloud samples of interchange structural units are obtained. Each complete point cloud sample of interchange structural units is processed to simulate missing regions, generating incomplete point cloud samples of structural units with simulated missing regions. The point cloud samples of interchange structural units refer to a pre-constructed set of point cloud data of various types of structural units that are complete in geometric shape, including complete instance point clouds of bridge deck, ramps, piers and guardrails, and each sample has corresponding semantic category labels and instance labels. Using incomplete structural unit point cloud samples as input, corresponding complete overpass structural unit point cloud samples as supervision labels, and the spatial position and geometric shape of adjacent structural unit nodes in the simulated missing region in the overpass structural topology map as constraints, the conditional generative point cloud completion model is iteratively trained until the completion result output by the model meets the preset accuracy requirements.

[0014] Furthermore, after generating the complete structural unit point cloud of the occluded structural unit node, the following verification steps are also included: The generated complete structural unit point cloud is compared with the structural unit point cloud of the adjacent structural unit node at the spatial connection point. If the comparison result indicates that the geometric shape at the boundary is continuous and there is no spatial overlap, the completion result is confirmed to pass the verification. The geometric continuity refers to the fact that the average distance between the corresponding nearest neighbor pairs in the point clouds on both sides of the boundary is less than a first preset threshold, and the included angle of the tangential direction at the corresponding point is less than a second preset threshold; the spatial interlacing refers to the fact that the point clouds on both sides of the boundary penetrate each other, and the volume or number of point clouds in the penetrated area exceeds a third preset threshold. If the comparison results indicate that there are geometric discontinuities or spatial overlaps at the boundary, the boundary geometry of the adjacent structural unit nodes will be used as a mandatory constraint to correct the boundary of the completion result until the completion result passes the verification.

[0015] Furthermore, the preset parametric structure template library contains parametric geometric templates corresponding to various structural unit types, including: straight bridge deck templates, curved ramp templates, bifurcation structure templates, and intersection structure templates; each parametric geometric template corresponds to a structural unit type, and the geometric shape of the structural unit type is defined by a set of adjustable geometric control parameters; The specific matching method is as follows: calculate the geometric feature similarity between the complete structural unit point cloud and the calibration geometric shape defined by each parameterized geometric template, and select the parameterized geometric template with the highest geometric feature similarity as the matching result; the geometric feature similarity includes the overall orientation similarity, width distribution similarity and elevation change similarity of the structural unit, and the similarity of the three dimensions is weighted and combined to obtain the final similarity score; Based on the geometric control parameters defined by the parameterized geometric template selected according to the matching results, the corresponding geometric feature values ​​are extracted from the point cloud of the complete structural unit, and the extracted geometric feature values ​​are used as the geometric control parameters of the structural unit node; the geometric feature values ​​include the centerline curvature radius, cross-sectional width, elevation change and end face tilt angle of the structural unit; The specific assembly method is as follows: Based on the adjacent structural unit nodes connected by each connected edge in the overpass structure topology diagram, determine the common connection end face between adjacent parametric 3D models; using the spatial position of each common connection end face as the alignment reference, perform spatial transformation and positioning on each parametric 3D model to achieve the alignment and splicing of adjacent end faces; After forming a complete three-dimensional semantic model of the overpass, the following verification steps are also included: traversing each intersection edge in the topology diagram of the overpass structure, checking whether there is spatial penetration or interference in the vertical direction of each structural unit model in the complete three-dimensional semantic model of the overpass; if there is spatial penetration or interference, the bottom surface of the upper structural unit model connected by the intersection edge is used as a constraint to avoid and correct the top surface of the lower structural unit model.

[0016] The present invention also provides a system for constructing a three-dimensional model of an overpass, which is used to implement the above-mentioned method for constructing a three-dimensional model of an overpass, including: The point cloud segmentation and clustering module is used to perform semantic segmentation and instance segmentation on the original point cloud data, resulting in multiple structural unit point clouds with semantic category labels and instance labels; The topology graph construction module is used to treat each structural unit point cloud as a structural unit node based on the instance label, establish the topological connection relationship between each structural unit node based on the spatial position of each structural unit node, and construct the topology graph of the overpass structure. The missing occlusion determination module is used to determine the standard reference geometric information corresponding to each structural unit node based on the overpass structure topology map, compare each structural unit point cloud with the corresponding standard reference geometric information, identify the regions in which each structural unit point cloud is missing relative to the corresponding standard reference geometric information, and determine the occluded structural unit node and its occlusion type based on the spatial position of the region in the corresponding structural unit point cloud and the topological connection relationship of the corresponding structural unit node in the overpass structure topology map. The neighborhood constraint completion module is used to take the incomplete structural unit point cloud of the occluded structural unit node as input according to the occlusion type, and take the spatial position and geometric shape of each structural unit node adjacent to the occluded structural unit node in the overpass structural topology map as constraints. It is then input into the pre-trained conditional generative point cloud completion model for inference to generate the complete structural unit point cloud of the occluded structural unit node. The template matching and assembly module is used to match the complete structural unit point cloud with templates in the preset parametric structural template library, select the template with the highest matching degree and fit the corresponding geometric control parameters to generate parametric 3D models of each structural unit node; and assemble each parametric 3D model according to the corresponding spatial position according to the topological connection relationship to form a complete 3D semantic model of the overpass.

[0017] Compared with the prior art, the technical effects and advantages provided by the present invention are as follows: This invention utilizes a deep learning model for end-to-end semantic and instance segmentation of airborne LiDAR point clouds, overcoming the shortcomings of traditional methods that rely on manually designed rules and parameter settings. It can adapt to point cloud data with varying densities and noise levels, significantly improving the accuracy and automation of overpass structural unit segmentation. Secondly, by constructing an overpass structural topology graph and introducing connected and intersecting edges to express the convergence, branching, and overlapping relationships between structural units, and combining the semantic category and topological context information of each structural unit to determine its standard reference geometry, the identification of occlusion areas is upgraded from traditional local endpoint detection to comprehensive detection based on global topological reasoning. This accurately identifies occlusion occurring at any location within a structural unit and determines the specific occlusion type. Furthermore, a differentiated generative completion strategy is adopted for different occlusion types. Compared to traditional repair methods that rely solely on one-dimensional interpolation of the centerline, this effectively restores the complete geometric shape of the occluded area, significantly improving the accuracy and model completeness of occlusion repair. Finally, parametric template matching generates a 3D model with semantic category labels and topological connectivity attributes, ensuring that the final output 3D semantic model of the overpass possesses both geometric accuracy and semantic richness. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a statistical summary diagram of the recall rates of the two methods under different occlusion types in this invention; Figure 3 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example: Please see Figures 1 to 2 The present invention provides a technical solution: A method for constructing a 3D model of an overpass, comprising the following steps: Step 1: Perform semantic segmentation and instance segmentation on the original point cloud data to obtain multiple structural unit point clouds with semantic category labels and instance labels.

[0022] In this embodiment, the original point cloud data is airborne lidar point cloud data; the semantic category labels include bridge deck category, ramp category, pier category, and guardrail category.

[0023] The raw point cloud data is preprocessed, including removing isolated outliers and normalizing coordinates; each point must contain at least three-dimensional coordinates and may include laser reflection intensity. Semantic category labels are encoded using integers, for example, bridge deck category is 0, ramp category is 1, pier category is 2, and guardrail category is 3. Instance labels are numbered sequentially from 1 for different connected components within the same semantic category; instance numbers can be repeated between different semantic categories.

[0024] The semantic segmentation and instance segmentation are achieved through a pre-trained 3D point cloud segmentation network. The 3D point cloud segmentation network extracts and classifies the semantic features of the original point cloud data point by point, and performs instance clustering based on the classification results and the spatial connectivity between point clouds, so as to distinguish the geometrical differences between unconnected point clouds under the same semantic category into different instances.

[0025] The 3D point cloud segmentation network employs a point-based deep learning segmentation network, whose structure includes an encoding part and a decoding part. The encoding part consists of five sequentially connected local feature aggregation modules and four downsampling layers, alternating between the two. The decoding part consists of four sequentially connected upsampling layers, four feature fusion layers, and one output layer. Each local feature aggregation module is used to extract local geometric features of the point cloud, and it includes a local spatial encoding unit, an attention pooling unit, and an extended residual unit. The local spatial encoding unit calculates the relative coordinates and Euclidean distance of each point's nearest neighbor, and concatenates the relative coordinates and distance to the original features of that neighbor, obtaining enhanced neighbor features. The attention pooling unit calculates attention weights for the enhanced neighbor features using a shared multilayer perceptron, and then performs a weighted summation of all neighbor features to obtain the aggregated features of that point. The extended residual unit performs residual concatenation between the aggregated features and the input features, and then performs feature transformation through two cascaded multilayer perceptrons. Each layer of the multilayer perceptron is followed by batch normalization and a ReLU activation function.

[0026] In this embodiment, the output feature dimensions of each local feature aggregation module in the encoding part are set to 16, 32, 64, 128, and 256 respectively; the number of nearest neighbor points in each local feature aggregation module is set to 16. Each downsampling layer adopts a random sampling method, randomly retaining 25% of the input points each time, so as to gradually reduce the point cloud resolution and expand the receptive field.

[0027] In the decoding section, each upsampling layer uses nearest-neighbor interpolation to restore the low-resolution feature map to the next higher resolution. Each feature fusion layer concatenates the upsampled features with the output features from the corresponding local feature aggregation module in the encoding section, and then fuses them through a feature fusion unit containing two multilayer perceptrons. The fused feature dimension is consistent with the feature dimension of the corresponding encoding layer. The output layer is a fully connected layer whose input is the feature output from the last feature fusion layer. The output dimension is 4, corresponding to the predicted probabilities of four semantic categories: bridge surface, ramp, pier, and guardrail. The category with the highest probability for each point is taken as the semantic classification result for that point.

[0028] The network is trained using manually labeled overpass point cloud samples as supervision, employing an end-to-end training method with cross-entropy loss. The optimizer is Adam, with an initial learning rate of 0.01 and a batch size of 8, training continues until the loss function converges. These parameters can be adjusted according to the point cloud size and available memory.

[0029] The instance clustering module takes semantic category labels and 3D coordinates as input. It uses density-based connectivity clustering to group spatially continuous points within the same semantic category into the same instance, assigning different instance labels to different connected components. This instance clustering module is executed as a post-processing step after the 3D point cloud segmentation network inference is completed.

[0030] The semantic segmentation and instance segmentation specifically include: first, performing point-by-point semantic classification on the original point cloud data and assigning a corresponding semantic category label to each point cloud data point. Then, for the point cloud data that has been assigned semantic category labels, performing density-based connectivity clustering based on the spatial location of each point cloud data point, dividing the spatially continuous point cloud set belonging to the same semantic category into an independent structural unit instance and assigning a corresponding instance label.

[0031] Density-based connectivity clustering uses three-dimensional Euclidean distance as the distance metric between points. The clustering algorithm can be DBSCAN, with a neighborhood radius of 3 times the average nearest neighbor distance of the point cloud and a minimum number of points of 20. The rationale for using 3 times the average nearest neighbor distance is that the density of airborne LiDAR point clouds varies in areas such as bridge surfaces and ramps; 3 times can bridge local fluctuations in point spacing and maintain the spatial continuity of the point cloud within the same structural unit. The minimum number of points of 20 can effectively suppress the misclassification of a small number of isolated points caused by noise or scan gaps as independent instances. The general method for determining the average nearest neighbor distance is as follows: first, calculate the distance from each point in the point cloud of each semantic category to its nearest neighbor, and take the average as the average nearest neighbor distance; the neighborhood radius is selected within the range of 2 to 5 times the average nearest neighbor distance, and the minimum number of points is selected within the range of 10 to 50, adjusted based on the criterion that the same structural unit is not broken and different structural units are not merged in the instance segmentation results. Clustering is performed only within the same semantic category label; density connectivity relationships are not established between different semantic categories; each connected component obtained from clustering is assigned a unique instance label.

[0032] Step 2: Based on the instance labels, treat each structural unit point cloud as a structural unit node, establish the topological connection relationship between each structural unit node according to the spatial position of each structural unit node, and construct the overpass structure topology map.

[0033] Each structural unit node directly corresponds to all point clouds with the same instance label in step 1; the spatial position of a node in the topology graph is represented by the horizontal projection centroid of the node's point cloud or the center of the axis-aligned bounding box. The topology graph of the overpass structure is represented by an undirected graph, with the node set consisting of all structural unit nodes and the edge set consisting of connected edges and intersecting edges.

[0034] In this embodiment, the topological connection relationship includes connected edges representing structural intersections and forks, and intersecting edges representing overlapping upper and lower spatial layers; the structural intersection refers to the topological connection relationship in which multiple structural unit nodes converge to one structural unit node along the driving direction, and the fork refers to the topological connection relationship in which one structural unit node branches into multiple structural unit nodes along the driving direction.

[0035] Connecting edges are used to express the physical connection between structural units within the same level, while intersecting edges are used to express the spatial spanning relationship between different levels. Intersections and bifurcations are represented in the diagram as a node having multiple incoming edges and one outgoing edge, or one incoming edge and multiple outgoing edges, respectively. They can be distinguished by counting the number of connected edges connected to the node and the direction of travel.

[0036] The method for establishing the connecting edge is as follows: when the projections of two structural unit nodes in space overlap, and the structural unit point clouds corresponding to the two structural unit nodes in the overlapping area are connected or intersecting in spatial position, a connecting edge is established between the two structural unit nodes.

[0037] The spatial projection uses a horizontal plane, specifically the XY plane. The overlapping projection area is determined by calculating the axis-aligned projection rectangles of the two node point clouds on the XY plane and judging whether the rectangles intersect. If the area of ​​the intersecting rectangle is greater than 5% of the area of ​​either projection rectangle, an overlapping projection area is considered to exist. Within the overlapping area, the Euclidean distance between all nearest neighbor pairs in the two node point clouds is calculated. If the minimum distance is less than a first distance threshold, they are considered connected or intersecting, and a connected edge is established. The first distance threshold is twice the average nearest neighbor distance of the interchange point cloud, typically 0.3 to 0.8 meters. The rationale for using twice the average nearest neighbor distance is that there are local gaps in the point clouds at the connection points of adjacent structural units; twice the average can accommodate scanning noise and slight gaps, while avoiding misjudging distant non-connected structures as connected. Generally, the distance between each point in the entire interchange point cloud and its nearest neighbor is calculated and averaged. Twice this average is used as the initial threshold. If obvious erroneous connections or missing connections appear after establishing connected edges, the threshold is adjusted within the range of 0.5 to 5 times the average.

[0038] The method for establishing the cross edge is as follows: when the projections of two structural unit nodes in space overlap, and the point clouds of the structural unit nodes corresponding to the two structural unit nodes in the overlapping area have vertical spacing and do not intersect each other, a cross edge is established between the two structural unit nodes.

[0039] The determination of overlapping projection areas is consistent with that of connected edges, using horizontally aligned rectangles that intersect and whose intersection area is greater than 5% of the area of ​​any projection. The vertical spacing is calculated as follows: within the overlapping area, take the elevation values ​​of all points in the point clouds of two nodes, calculate the difference between the maximum elevation of one node's point cloud and the minimum elevation of the other's point cloud, and take the smaller of the two differences as the vertical clearance. If this vertical clearance is greater than a second distance threshold and the nodes do not intersect, then a crossing edge is established. The second distance threshold is 2 meters, with an upper limit of 20 meters. The rationale for 2 meters is that the design clearance between adjacent layers of an overpass is generally not less than 2 meters; anything less can be considered undulation or interference within the same layer. The upper limit of 20 meters can eliminate the possibility of piers and bridge decks being mistakenly identified as crossing edges, because piers and bridge decks are connected rather than intersecting. Generally, the spacing is selected based on the overpass design clearance standard or the layer spacing distribution in the point cloud height histogram, typically ranging from 1.5 to 5 meters; if there are special low-rise or super high-rise buildings, the upper limit can be adjusted. Non-intersecting means that the three-dimensional nearest neighbor distance between two node point clouds within the overlapping region is greater than a first distance threshold. The two nodes connected by intersecting edges both store upper and lower layer identifiers in the graph, where the minimum elevation of the upper layer node point cloud is greater than the maximum elevation of the lower layer node point cloud.

[0040] Each structural unit node is associated with and stores the following attribute information: the spatial bounding box of the structural unit point cloud corresponding to the structural unit node, the structural orientation description information, the elevation description information, and the corresponding semantic category label.

[0041] The spatial bounding box adopts the axis-aligned minimum bounding box, that is, the minimum and maximum values ​​of the point cloud in the X, Y, and Z directions are calculated respectively to obtain the bounding box range represented by six coordinate values; this bounding box is also used for determining the projection overlap area and subsequent spatial position constraints. The semantic category label adopts the integer encoding of step 1. The structural orientation description information and elevation description information are extracted according to semantic categories, and the specific methods are described in subsequent paragraphs.

[0042] When the semantic category label is bridge deck or ramp, the structural orientation description information includes centerline orientation or trajectory line orientation. For bridge decks and ramps, the steps for extracting the centerline orientation are as follows: project the point cloud onto the XY plane and remove the Z coordinate; perform two-dimensional principal component analysis on the projected points and take the direction with the largest variance as the main orientation; divide the main orientation into strips at fixed intervals, such as 1 meter, and calculate the centroid of the projected points in each strip to obtain a series of discrete center points; then perform cubic B-spline curve fitting or polygonal line connection on the discrete center points to obtain the centerline. The centerline is stored as a three-dimensional coordinate sequence arranged in ascending order of mileage, and the elevation of each point is taken as the average elevation of the point cloud within that strip. If the ramp has branches or intersections, the centerlines of each branch can be extracted segment by segment.

[0043] When the semantic category label is "pier," the structural orientation description information includes the vertical axis direction. Pier point clouds are typically slender columnar, and the vertical axis direction is obtained through three-dimensional principal component analysis of the pier point cloud, taking the eigenvector corresponding to the largest eigenvalue as the axis direction; if the pier is tilted, this vector reflects the actual tilt direction. Simultaneously, a reference point along the axis is stored, which can be the three-dimensional centroid of the pier point cloud.

[0044] When the semantic category label is "guardrail category", the structural orientation description information includes the horizontal trajectory orientation. The guardrail is a horizontal strip structure extending along the edge of the bridge deck. The extraction method of its horizontal trajectory orientation is similar to that of the bridge deck centerline: the guardrail point cloud is projected onto the XY plane, and the projected points are fitted with curves or connected by polylines to obtain the horizontal trajectory line; it is stored as a two-dimensional coordinate sequence arranged by mileage, and the elevation is taken as the average elevation of the point cloud within the corresponding strip.

[0045] The elevation description information includes elevation range or centerline elevation profile data. The elevation range refers to the minimum and maximum values ​​of the point cloud in the Z direction; the centerline elevation profile data is a sequence of elevation values ​​recorded at fixed mileage intervals along the centerline or trajectory line, used to describe the vertical changes of structural units. For bridge piers, the elevation range from bottom to top along the vertical axis can be recorded; for guardrails, the elevation of each point on the horizontal trajectory can be recorded.

[0046] Step 3: Determine the standard reference geometric information corresponding to each structural unit node based on the overpass structure topology map, compare each structural unit point cloud with the corresponding standard reference geometric information, and identify the areas where each structural unit point cloud is missing relative to the corresponding standard reference geometric information; based on the spatial position of the area in the corresponding structural unit point cloud and the topological connection relationship of the corresponding structural unit node in the overpass structure topology map, determine the occluded structural unit node and its occlusion type.

[0047] In this embodiment, the standard reference geometry information is determined based on the semantic category labels of the structural unit nodes: When the semantic category label of a structural unit node is a bridge deck or ramp, the standard reference geometric information includes the reference centerline orientation, reference width variation range, and reference elevation variation range corresponding to the structural unit. Specifically: based on the centerline extracted in step 2, samples are taken along the centerline at fixed intervals, such as 1 meter. At each sampling location, the lateral distance distribution from the point cloud to the centerline in the neighborhood of that location is statistically analyzed. The 95th percentile of the lateral distance is taken as the reference half-width at that location, thus obtaining the reference width variation range. At the same time, the median of the elevation of the point cloud in the neighborhood is taken as the reference elevation at that location, forming the reference elevation variation range.

[0048] When the semantic category label is a bridge pier, the standard reference geometric information includes the reference vertical axial orientation and the reference cross-sectional size range corresponding to the structural unit; specifically: for the bridge pier category, with the vertical axis as the center, the maximum distance from the point cloud in the cross-section perpendicular to the axis to the axis is counted, and the median of the maximum distances of all cross-sections is taken as the reference cross-sectional radius to obtain the reference cross-sectional size range.

[0049] When the semantic category label is a guardrail category, the standard reference geometric information includes the reference horizontal trajectory direction and reference height profile range corresponding to the structural unit. Specifically: based on the horizontal trajectory line, the distribution range of the point cloud in the vertical direction is statistically analyzed along the trajectory line, and the median of the height range is taken as the reference height profile.

[0050] The reference centerline and reference horizontal trajectory are directly adopted from the coordinate sequence of the centerline or trajectory stored in step 2; the reference vertical axis is directly adopted from the vertical axis direction stored in step 2. The above statistics use the median or high quantile instead of the maximum value to avoid interference from extreme values ​​caused by individual outliers or local occlusion, so that the standard reference geometric information reflects the main geometric shape of the structural unit.

[0051] For each point cloud data point in each structural unit point cloud, calculate the degree of spatial deviation between its spatial position and the standard position corresponding to the standard reference geometric information; identify the continuous region formed by point cloud data points whose spatial deviation exceeds the preset tolerance range as a missing region.

[0052] The specific process for calculating the spatial deviation of each point cloud data point is as follows: First, determine the corresponding position of the point in the standard reference geometry. For bridge decks or ramps, project the point onto the reference centerline, find the nearest centerline mileage position, and calculate the lateral horizontal distance between the point and the reference centerline at that position, as the lateral deviation. Calculate the difference between the point's elevation and the reference elevation at that position, as the elevation deviation. If the lateral deviation exceeds 1.1 times the reference half-width, or the absolute value of the elevation deviation exceeds 0.5 meters, the point is marked as a deviation point. For bridge piers, calculate the radial distance from the point to the vertical axis. If the radial distance exceeds 1.2 times the reference cross-sectional radius, it is marked as a deviation point. For guardrails, calculate the horizontal and vertical distances from the point to the horizontal trajectory line. If the horizontal distance exceeds 1.2 times the corresponding horizontal range in the reference height profile, or the vertical distance exceeds the upper or lower limit of the height profile by more than 0.3 meters, it is marked as a deviation point. Among the aforementioned threshold values, 1.1x, 1.2x, 0.5m, and 0.3m are empirical values. Their rationale is to allow for normal point cloud noise and minor geometric fluctuations, but anything exceeding 10% to 20% is considered an abnormal loss. The general method for determining these thresholds is as follows: first, statistically analyze the distance deviation distribution from all points in the point cloud of each structural unit to their corresponding standard positions, and take the 90th to 95th percentile of this deviation distribution as the upper tolerance limit; if the identified missing area is significantly larger or smaller than expected, adjust this percentile accordingly.

[0053] All point cloud data points marked as deviation points are clustered using the same density-based connectivity clustering method as in step 1. The neighborhood radius is set to twice the average nearest neighbor distance of the point cloud, and the minimum number of points is 10. Each connected component obtained from the clustering is considered a candidate missing region. If the number of points in a candidate missing region is less than 1% of the total number of points in the point cloud of that structural unit, it is discarded to avoid interference from noise in determining the occlusion type of small regions.

[0054] In this embodiment, the occlusion types include: When the spatial location of the missing region in the corresponding structural unit point cloud completely overlaps with the spatial projection area of ​​any structural unit node connected by an intersecting edge in the overpass structural topology map, it is determined to be a full occlusion type.

[0055] When the missing region and the spatial projection area of ​​at least one structural unit node connected by intersecting edges in the overpass structure topology map only partially overlap, and there is no complete overlap with the spatial projection area of ​​any structural unit node, it is determined to be a partial occlusion type; the spatial projection area is the vertical projection coverage of the structural unit point cloud corresponding to each structural unit node on the horizontal plane, and its determination is based on the projection range of the spatial bounding box stored by each structural unit node on the horizontal plane.

[0056] The spatial projection area of ​​a missing region is represented by the convex hull or axis-aligned rectangle of the horizontal projections of all point cloud points within that region. Complete overlap of spatial projection areas means that the area overlapping between the projected area of ​​the missing region and the projected rectangles of the structural unit nodes connected by the intersecting edges accounts for more than 90% of the projected area of ​​the missing region. Partial overlap means that the overlapping area accounts for 10% to 90% of the projected area of ​​the missing region, and there is no situation where the projected rectangle of any node connected by the intersecting edge covers more than 90% of the projected area of ​​the missing region. The rationale for setting the ratio thresholds to 90% and 10% is that it can distinguish between completely covered and partially covered missing regions, while avoiding misjudgments caused by projection boundary alignment errors. The general method for determining this ratio is as follows: first, based on a visual comparison between the projected rectangles of the actual intersecting edge nodes and the projection of the missing region, select an initial ratio threshold, such as 80%, and then adjust it according to the judgment results.

[0057] When the continuity of the connected edges of an obscured structural unit node in the overpass structural topology diagram is interrupted by a missing region, it is determined to be a structural break obscuration type. The determination method for the interruption of the continuity of the connected edges by a missing region is as follows: when the missing region divides the corresponding structural unit point cloud into at least two spatially unconnected point cloud segments along the driving direction, and the upstream and downstream structural unit nodes connected by the obscured structural unit node in the overpass structural topology diagram through connected edges are located exactly within these two point cloud segments, it is determined that the continuity of the connected edges is interrupted by a missing region.

[0058] For full occlusion, the missing region spans the entire width of the corresponding structural unit's point cloud. Specifically, the missing region's maximum lateral span in the direction perpendicular to the reference centerline is greater than or equal to 90% of the reference half-width at that location. For partial occlusion, if the missing region's maximum lateral span is less than 90% of the reference half-width and its length along the direction is less than 50% of the total length of the structural unit, then it only covers a portion of the width or length.

[0059] For structural break occlusion types, the occluded structural unit node is connected to upstream and downstream structural unit nodes respectively through connecting edges in the overpass structural topology diagram. When determining the structural break occlusion type, first check whether the missing region divides the point cloud into two unconnected point cloud segments along the reference centerline direction of the structural unit node; then determine the structural unit nodes connected by the upstream and downstream connecting edges along the driving direction. If the upstream node's point cloud is located within one point cloud segment and the downstream node's point cloud is located within another point cloud segment, then the continuity of the connecting edge is determined to be interrupted by the missing region.

[0060] The missing region divides the point cloud of the structural unit corresponding to the occluded structural unit node into at least two spatially unconnected point cloud segments.

[0061] Step 4: Based on the occlusion type, the incomplete structural unit point cloud of the occluded structural unit node is taken as input. The spatial position and geometric shape of each structural unit node adjacent to the occluded structural unit node in the overpass structural topology map are used as constraints. The inference is performed in the pre-trained conditional generative point cloud completion model to generate the complete structural unit point cloud of the occluded structural unit node.

[0062] In this embodiment, the spatial location includes the spatial coordinates of each adjacent structural unit node, and the geometric shape includes the structural unit width, structural unit orientation vector, and corresponding semantic category label of each adjacent structural unit node.

[0063] Spatial location coordinates are taken from the 3D centroid coordinates of the point clouds of adjacent structural unit nodes, or from the coordinates of the eight corner points of their spatial bounding box. The structural unit width is taken as the average cross-sectional width of the node in the direction perpendicular to its orientation, and the structural unit orientation vector is taken as the tangential unit vector of the node's centerline at the connecting end face. The semantic category label is the integer code from step 1. This information is normalized and concatenated into a fixed-length conditional vector, which is then input into the model along with the point cloud data.

[0064] One specific implementation of the conditional generative point cloud completion model adopts a conditional generative adversarial network architecture, specifically including: The incomplete structural unit point cloud and the point clouds of adjacent structural unit nodes are resampled uniformly to ensure a fixed total number of points in the input point cloud of 2048. When an occluded node has its own unmissing point cloud, it is resampled to 1024 points; the point clouds of adjacent nodes are also resampled to 1024 points after being stitched together. Each point is input with 3D coordinates, and under the partial occlusion type, a binary mask indicating whether the point belongs to a missing region is also input, so the actual input dimension of each point is 4. The condition vector is obtained by encoding the structural unit width, structural unit orientation vector, and semantic category label of adjacent nodes, with a dimension of 64.

[0065] The generator adopts a PointNet++ encoder-decoder architecture: the encoder contains four ensemble abstraction layers; the first ensemble abstraction layer samples 512 points, with a ball query radius of 0.2 meters, and samples 32 points per group, with the number of point-level multilayer perceptron channels [64, 64, 128], and the output feature dimension is 128; the second ensemble abstraction layer samples 128 points, with a ball query radius of 0.4 meters, and samples 32 points per group, with the number of point-level multilayer perceptron channels [128, 128, 256], and the output feature dimension is 256; the third ensemble abstraction layer samples 32 points, with a ball query radius of 0.8 meters, and samples 16 points per group, with the number of point-level multilayer perceptron channels [256, 256, 512], and the output feature dimension is 512; the fourth ensemble abstraction layer samples 1 point, with a ball query radius of 2.0 meters, groups all remaining points, and has the number of point-level multilayer perceptron channels [512, 512, 1024], outputting a global feature dimension of 1024. Sampling radii of 0.2 m, 0.4 m, 0.8 m, and 2.0 m correspond to local, detailed, cross-unit, and global ranges in the point cloud of the overpass, respectively. This allows for a progressively larger receptive field, covering details within structural units and spatial relationships between adjacent units.

[0066] The 1024-dimensional global features are concatenated with the 64-dimensional conditional vector to obtain 1088-dimensional features. The decoder consists of three fully connected layers: the first fully connected layer takes 1088-dimensional input and outputs 1024-dimensional features; the second fully connected layer takes 1024-dimensional input and outputs 1024-dimensional features; and the third fully connected layer takes 1024-dimensional input and outputs 6144-dimensional features, reshaping the data into a complete 2048×3 point cloud coordinate system. Each fully connected layer is followed by batch normalization and a ReLU activation function, while the last layer does not have an activation function.

[0067] The discriminator uses a PointNet architecture; the input is either the complete point cloud output by the generator or a real complete point cloud, with a fixed number of points of 2048. It first passes through three shared-point multilayer perceptrons with channel numbers of [64, 128, 256], followed by batch normalization and a LeakyReLU activation function at each layer. Then, max pooling is performed on the 256-dimensional features to obtain 256-dimensional global features. This global feature is concatenated with a 64-dimensional conditional vector, and then passed through two fully connected layers. The first layer takes a 320-dimensional input and outputs a 256-dimensional vector, while the second layer takes a 256-dimensional input and outputs a 1-dimensional vector. Finally, a Sigmoid activation is applied to output the true / false probability.

[0068] The optimizer uses Adam with an initial learning rate of 0.001, a batch size of 16, and 200 training epochs. The learning rate decays to 0.5 every 50 epochs. The generator and discriminator are updated alternately, with the generator updating after the discriminator. The sampling radius, number of points, and number of channels can be adjusted according to the actual point cloud density and structural unit size. A common method is to calculate the average nearest neighbor distance of the point cloud, taking the radius of the first abstraction layer as 5 to 10 times the average nearest neighbor distance, doubling the radius layer by layer thereafter; the number of points can be reduced to 1 / 4, 1 / 16, or 1 / 64 of the total input points; the number of output points is twice the number of input points to preserve the surface details of the completed result.

[0069] The conditional generative point cloud completion model employs differentiated completion strategies based on different occlusion types, specifically including: When the occlusion type is full occlusion, the conditional generative point cloud completion model uses the spatial position and geometric shape of adjacent structural unit nodes as the only morphological reference to generate the complete structural unit point cloud of the occluded structural unit node. When the occlusion type is partial occlusion, the conditional generative point cloud completion model uses the geometry of the missing part in the incomplete structural unit point cloud of the occluded structural unit node as a reference, and combines the spatial position and geometry of the adjacent structural unit node to perform local completion and generate a complete structural unit point cloud. When the occlusion type is structural fracture occlusion, the conditional generative point cloud completion model takes the point cloud segments on both sides of the fracture of the occluded structural unit node as the reference, and the spatial position and geometry of the adjacent structural unit node as the guide to generate the intermediate transition segment point cloud connecting the two side point cloud segments.

[0070] In this embodiment, the specific implementation of the differentiated completion strategy is as follows: Under the full occlusion type, the model input only includes the point cloud and condition vector of the adjacent nodes, without inputting the point cloud of the occluded node itself, and the generator directly outputs the complete point cloud; Under the partial occlusion type, the point cloud of the unmissing part is concatenated with the point cloud of the adjacent node as input, the model retains the unmissing part and generates in the missing area, and the position of the missing area can be marked by a mask; Under the structural fracture occlusion type, the point cloud segments on both sides of the fracture are used as anchor points, and the point cloud segments on both sides and the point cloud of the adjacent node are input into the model, the model generates the intermediate transition segment point cloud connecting the two sides, and concatenates the point cloud segments on both sides with the generated segment to form a complete point cloud.

[0071] In this embodiment, the pre-training method of the conditional generative point cloud completion model is as follows: Multiple complete point cloud samples of interchange structural units are acquired. Simulated missing regions are then processed on each complete point cloud sample to generate incomplete point cloud samples with simulated missing areas. These point cloud samples refer to a pre-constructed set of point cloud data for various types of structural units that are geometrically complete, including complete instance point clouds of bridge decks, ramps, piers, and guardrails. Each sample carries corresponding semantic category and instance labels. A geometrically complete point cloud data set means that the structural unit point cloud does not contain holes, breaks, or edge defects caused by occlusion, scan loss, or noise; the point cloud density is uniform; and it completely covers the actual geometric surface of the structural unit. The specific acquisition methods include the following two: Method 1: Multi-source point cloud fusion and manual repair; For the same overpass scene, airborne LiDAR point clouds, ground-based laser scanning point clouds, and oblique photogrammetry point clouds are simultaneously collected. The three types of point clouds are registered and fused using an iterative nearest-point algorithm. After fusion, the point cloud density is increased, and occluded areas are complemented. For local holes and missing edges still existing in the fused point cloud, the radial basis function implicit surface reconstruction method is used to fill the holes, and then resampling is performed according to the original point cloud density. Then, the repaired point cloud is manually semantically annotated instance by instance, assigning a unique semantic category label and instance label to each structural unit, resulting in a complete structural unit point cloud sample.

[0072] Method 2: Generating a simulated point cloud based on a parametric structural template; using the parametric structural template library described in step 5, multiple sets of different geometric control parameters are set for each structural unit type, such as a bridge deck width of 8 to 16 meters, a ramp curvature radius of 30 to 120 meters, a pier cross-sectional radius of 0.6 to 1.5 meters, and a guardrail height of 0.8 to 1.2 meters. Based on the geometric shape defined by the parametric geometric template, Poisson disk sampling is performed on the structural surface at a preset point density to generate a simulated point cloud with complete geometric shape and controllable density, and the corresponding semantic category label and instance label are automatically assigned. This method can be used in combination with Method 1 to expand the number and diversity of samples.

[0073] Using incomplete structural unit point cloud samples as input, corresponding complete overpass structural unit point cloud samples as supervision labels, and the spatial position and geometric shape of adjacent structural unit nodes in the simulated missing region in the overpass structural topology map as constraints, the conditional generative point cloud completion model is iteratively trained until the completion result output by the model meets the preset accuracy requirements.

[0074] The simulated missing data processing is constructed according to three occlusion types. Full occlusion samples are generated by randomly selecting a complete structural unit point cloud, deleting the entire point cloud segment, and retaining only adjacent node information. Partial occlusion samples are generated by randomly cropping a locally connected region of the structural unit point cloud; the cropped region can be located inside or at the edge of the point cloud, allowing the remaining point cloud to coexist with the missing region. Structural fracture occlusion samples are generated by deleting a segment of the structural unit point cloud along its direction, dividing the remaining point cloud into two segments. When simulating missing data, the adjacent structural unit node information corresponding to the missing region is retained as a constraint. The training loss function is the sum of the chamfer distance and the bulldozer distance between the generated point cloud and the real complete point cloud. The chamfer distance calculation formula is: in, This is the chamfer distance. To generate point clouds, For realistic point clouds, Indicates the generation of point clouds The index of the point in, Representing real point clouds The index of the point in; Point With point The three-dimensional Euclidean distance between them.

[0075] The bulldozer distance was approximated using the Sinkhorn algorithm. The model accuracy was considered satisfactory when the average chamfer distance on the validation set was less than 0.03 meters and the average bulldozer distance was less than 0.05 meters. The general method for determining the distance threshold was based on the criterion of visually inspecting the completed parts for obvious shape errors and natural boundary connections, combined with point cloud density.

[0076] In this embodiment, after generating the complete structural unit point cloud of the occluded structural unit node, the following verification steps are also included: The generated complete structural unit point cloud is compared with the structural unit point cloud of the adjacent structural unit node at the spatial connection point. If the comparison result indicates that the geometric shape at the boundary is continuous and there is no spatial overlap, the completion result is confirmed to pass the verification. The geometric continuity refers to the fact that the average distance between the corresponding nearest neighbor pairs in the point clouds on both sides of the boundary is less than a first preset threshold, and the included angle of the tangential direction at the corresponding point is less than a second preset threshold; the spatial interlacing refers to the fact that the point clouds on both sides of the boundary penetrate each other, and the volume or number of point clouds in the penetrated area exceeds a third preset threshold. If the comparison results indicate that there are geometric discontinuities or spatial overlaps at the boundary, the boundary geometry of the adjacent structural unit nodes will be used as a mandatory constraint to correct the boundary of the completion result until the completion result passes the verification.

[0077] In this embodiment, the first preset threshold is 1.5 times the average nearest neighbor distance of the overpass point cloud, typically 0.2 to 0.5 meters; the second preset threshold is 15 degrees; and the third preset threshold is 5% of the number of points in the penetrating area point cloud as a percentage of the total number of points in the boundary overlapping area. The rationale for these thresholds is that 1.5 times the average nearest neighbor distance can accommodate point cloud noise and slight boundary misalignment, 15 degrees maintains geometric continuity without allowing obvious angles, and the 5% penetration point ratio can distinguish between slight calculation errors and actual spatial overlap. A common method for determining these thresholds is to first statistically analyze the nearest neighbor distance distribution and angle distribution at the boundary on a verification sample with known correct completion results, taking the 90th percentile as the threshold, and then adjusting it according to the actual situation. Boundary correction is achieved by replacing the point cloud in the completion result that overlaps with the boundary of adjacent nodes with sampling points on the locally fitted surface of the point cloud of the adjacent node boundary, iteratively correcting until geometric continuity and no spatial overlap are satisfied.

[0078] Step 5: Match the complete structural unit point cloud with the templates in the preset parametric structural template library, select the template with the highest matching degree and fit the corresponding geometric control parameters to generate parametric 3D models of each structural unit node; and assemble each parametric 3D model according to its corresponding spatial position according to the topological connection relationship to form a complete 3D semantic model of the overpass.

[0079] Before matching, voxel downsampling is performed on the complete structural unit point cloud to unify the point spacing to 0.1 meters, thereby reducing the impact of point cloud density differences on similarity calculation. The downsampled point cloud is used for feature extraction and similarity calculation, while the original complete point cloud is retained for subsequent geometric control parameter extraction.

[0080] In this embodiment, the preset parametric structure template library contains parametric geometric templates corresponding to various structural unit types, including: straight bridge deck template, curved ramp template, bifurcation structure template and intersection structure template; each parametric geometric template corresponds to a structural unit type, and the geometric shape of the structural unit type is defined by a set of adjustable geometric control parameters.

[0081] The geometric control parameters for the straight bridge deck template include: bridge deck length, bridge deck width, bridge deck thickness, longitudinal slope, and transverse slope; the geometric control parameters for the curved ramp template include: centerline radius of curvature, central angle, ramp width, ramp thickness, superelevation, and longitudinal slope; the geometric control parameters for the bifurcation structure template include: mainline width, branch width, bifurcation angle, transition section length, and thickness at the bifurcation point; the geometric control parameters for the intersection structure template include: width of each intersection branch, intersection angle, transition section length, and thickness at the intersection point. All of the above parameters can be adjusted independently. The template generates a three-dimensional solid model by sweeping the cross-section along the centerline and applying thickness and slope.

[0082] The specific matching method is as follows: calculate the geometric feature similarity between the point cloud of the complete structural unit and the calibration geometric shape defined by each parameterized geometric template, and select the parameterized geometric template with the highest geometric feature similarity as the matching result; the geometric feature similarity includes the overall orientation similarity of the structural unit, the width distribution similarity, and the elevation change similarity, and the similarity of the three dimensions is weighted and combined to obtain the final similarity score.

[0083] Overall orientation similarity is obtained by comparing the Frescher distance between the centerline orientation of the complete structural unit point cloud and the reference centerline orientation of each template under standard parameters. The smaller the Frescher distance, the more similar the orientations. Width distribution similarity is obtained by comparing the normalized root mean square error between the cross-sectional width sequence of the point cloud at each position along the centerline and the calibration width sequence of the template at each position. Elevation variation similarity is obtained by comparing the normalized root mean square error between the elevation sequence of the point cloud along the centerline and the reference elevation sequence of the template. The three similarities are mapped to the interval 0 to 1, with closer to 1 indicating greater similarity. The final similarity score is calculated using the following weighted formula: in, To score the similarity, To ensure overall similarity, For width distribution similarity, The similarity is based on elevation changes. The rationale for weights of 0.5, 0.3, and 0.2 is as follows: Orientation is the most important feature distinguishing straight, curved, bifurcation, and intersection structures, thus receiving the highest weight; width distribution reflects cross-sectional changes and bifurcation characteristics, receiving the next highest weight; elevation changes are affected by slope and inter-layer height, contributing less to template differentiation, and therefore receiving the lowest weight. The general method is to select several known categories of structural unit point clouds from the template library as a validation set, adjust the three weights to maximize template matching accuracy, and keep the sum of the weights constant at 1.

[0084] Based on the geometric control parameters defined by the parameterized geometric template selected according to the matching results, the corresponding geometric feature values ​​are extracted from the point cloud of the complete structural unit, and the extracted geometric feature values ​​are used as the geometric control parameters of the structural unit node; the geometric feature values ​​include the centerline curvature radius, cross-sectional width, elevation change and end face tilt angle of the structural unit.

[0085] The specific extraction methods are as follows: For straight bridge deck templates, the horizontal projection length of the line connecting the beginning and end of the point cloud centerline is taken as the bridge deck length, the median of the widths of each section is taken as the bridge deck width, and the ratio of the elevation difference between the beginning and end to the horizontal length is taken as the longitudinal slope. For curved ramp templates, the centerline is first fitted with a circle, the radius of the fitted circle is taken as the radius of curvature, the angle between the tangential directions of the beginning and end of the centerline is taken as the central angle, and the median of the widths of each section is taken as the ramp width. For bifurcation structure templates, the horizontal angle between the centerline of each branch and the centerline of the main line is calculated at the bifurcation points of the main line and branches, the average value is taken as the bifurcation angle, the median of the widths of each section of the main line is taken as the main line width, and the median of the widths of each section of the branch is taken as the branch width. For intersection structure templates, the widths and intersection angles of each intersection branch are obtained using the same method. The end face tilt angle is obtained by calculating the angle between the normal vector of the tangent plane of the point cloud at the end face of the structural unit and the tangential direction of the centerline. All extracted values ​​are in meters or degrees and are rounded to two decimal places.

[0086] The specific assembly method is as follows: based on the adjacent structural unit nodes connected by each connecting edge in the overpass structure topology diagram, determine the common connection end face between adjacent parametric 3D models; using the spatial position of each common connection end face as the alignment reference, perform spatial transformation and positioning on each parametric 3D model to achieve alignment and splicing of adjacent end faces.

[0087] The method for determining the common connection end face is as follows: For two structural unit nodes connected by a connecting edge, find the end face that is closest to the center of the bounding box of the other model space on their respective parametric 3D models as the candidate connection end face; if the closest distance between the two candidate end faces is less than 0.5 meters, then the two are identified as the common connection end face.

[0088] Spatial transformation positioning employs rigid body transformation, including rotation and translation. Specifically, the normal vectors and centroid coordinates of the two common connecting end faces are first calculated. The normal vector of one model's connecting end face is rotated so that its direction is opposite to that of the normal vector of the other model's connecting end face. Then, the two end faces are translated so that their centroids coincide. The rotation matrix is ​​calculated using the Rodriguez rotation formula, where the rotation axis is the cross product of the two normal vectors, and the rotation angle is the angle between the two normal vectors. After the transformation, the two models are aligned at the common connecting end face. When multiple connected edges exist, the end face of the positioned model is used as the reference, and the transformations are performed sequentially on the unpositioned models.

[0089] After forming a complete three-dimensional semantic model of the overpass, the following verification steps are also included: traversing each intersection edge in the topology diagram of the overpass structure, checking whether there is spatial penetration or interference in the vertical direction of each structural unit model in the complete three-dimensional semantic model of the overpass; if there is spatial penetration or interference, the bottom surface of the upper structural unit model connected by the intersection edge is used as a constraint to avoid and correct the top surface of the lower structural unit model.

[0090] The method for checking for spatial penetration or interference is as follows: For each upper-level structural unit model and lower-level structural unit model connected by an intersecting edge, calculate their respective bounding boxes. If the vertical overlap height of the two bounding boxes is greater than 0.2 meters, then perform a Boolean intersection operation on the two solid models. If the intersection volume is greater than 0.1 cubic meters, or the number of point clouds in the intersection region is greater than 50, then spatial penetration or interference is determined to exist.

[0091] The rationale for setting thresholds of 0.2 meters, 0.1 cubic meters, and 50 points is as follows: vertical overlaps less than 0.2 meters may be due to approximation errors in the rounded corners or curved surfaces of the model ends, and do not constitute actual interference; 0.1 cubic meters and 50 points can filter out minor spurious intersections caused by numerical calculation errors. The general method for determining these thresholds is as follows: first, using a known correct model without interference as a sample, statistically analyze the maximum intersection volume and the number of point clouds caused by modeling errors, and then set the threshold to twice this statistical value.

[0092] The avoidance correction is achieved by modifying the elevation of the top surface of the lower structural unit model within the projection range of the intersection edge: the elevation value of the top surface of the lower model within this range is shifted downward as a whole, so that the top surface of the lower model is at least 0.1 meters lower than the bottom surface of the upper model, and the shifted top surface is smoothly transitioned to the adjacent unadjusted area to avoid generating new geometric discontinuities.

[0093] In this embodiment, the smooth transition between the lowered top surface and the adjacent unadjusted area specifically includes: Projecting the projection range of the upper-level structural unit model connected by the intersecting edges onto the top surface of the lower-level structural unit model on the horizontal plane yields the core area for avoidance correction; extending the boundary of the core area outwards forms a ring-shaped transition zone. The width of the transition zone... It can be based on the average width of the top surface of the lower structural unit. Average point spacing in point clouds Or the mesh size of the parameterized template is determined, for example, by taking... ,and No more than the width of the top surface of the lower structural unit 30%; 0.5m is the minimum allowable width of the transition zone, i.e., 0.5 meters. For bridge deck or ramp substructure units, the transition zone can be set longitudinally and laterally along the centerline.

[0094] Then, an elevation correction function is established on the top surface of the lower structural unit. Let any point on the top surface of the lower structural unit... The original elevation is The shortest distance from this point to the boundary of the core area is Within the core area, a downward displacement is applied to the top surface. The elevation of the lowered top surface It is at least 0.1m below the bottom surface of the upper structural unit; among which, The elevation difference between the bottom surface of the upper layer and the original top surface of the lower layer can be determined point by point, while meeting the clearance requirements.

[0095] Within the transition zone, a smoothing weighting function is used to gradually decay the downward shift from the core boundary towards the unadjusted region to zero. For example, let the normalized distance... The weighting function is: in, This represents the smoothing weight function within the transition band, with values ​​ranging from [0,1]. That is, when located at the boundary of the core area, ;when That is, when located at the outer boundary of the transition zone, The elevation correction for any point within the transition zone is as follows: in, For point Project along the normal or the nearest distance to the corresponding point on the boundary of the core area. This represents the downward displacement of the corresponding point. Therefore, the top surface of the core area is shifted downwards as a whole according to the required clearance. The downward displacement within the transition zone decreases smoothly until the elevation at the outer boundary of the transition zone matches the original top surface elevation of the adjacent unadjusted area. Furthermore, the first derivative of the elevation change is continuous, avoiding steps, angles, or new geometric discontinuities. In addition to the aforementioned triple smoothing weights, cosine weights can also be used. Equivalent smoothing can be achieved through methods such as B-spline interpolation or radial basis function interpolation.

[0096] For parametric 3D models, the elevation correction can be achieved by adjusting the elevation control points, cross-sectional elevation parameters, or swept cross-sectional control parameters of the top surface of the lower structural unit, and the parametric surface of the top surface can be regenerated. For triangular mesh or grid models, the elevation coordinates of each vertex in the transition zone can be updated according to the above weights, and the vertex normals can be recalculated, retriangulated, or local Laplacian smoothing can be performed.

[0097] After the smooth transition is completed, verification can be performed: check whether the height difference between adjacent vertices within the transition zone is less than a preset height difference threshold, such as 0.05m; check whether the rate of change of longitudinal and lateral slopes within the transition zone is less than a preset rate of change threshold, such as 5%; check whether the included angle of normal at the boundary of the transition zone is less than a preset angle threshold, such as 15°. If the thresholds are exceeded, the width of the transition zone is increased or the weighting function is readjusted until the lowered top surface is smoothly connected to the adjacent unadjusted area, and the spatial interference check of the intersection edge is re-executed.

[0098] This embodiment designed a set of comparative experiments to evaluate the performance advantage of the proposed method over traditional methods in the detection of occluded structures. The experiment used point cloud data of overpasses collected by an actual airborne LiDAR in a certain area as the test set. Thirty structural units exhibiting occlusion were randomly selected from this dataset as independent test samples, including 10 instances of full occlusion, 10 instances of partial occlusion, and 10 instances of structural fracture occlusion. Occlusion detection was performed on each sample using both the traditional local endpoint buffer detection method (hereinafter referred to as the traditional method) and the global topology inference-based detection method of this embodiment (hereinafter referred to as the method of this embodiment). The recall rates of the two methods under different occlusion types were statistically analyzed. To eliminate random errors, the detection results for each test sample were recorded separately, and the average and standard deviation of the recall rates under each occlusion type were calculated. The results are summarized in Table 1, as follows: Table 1: Comparison of experimental data on recall rates of the two methods under different occlusion types: Table 2: Summary of recall rates for the two methods under different occlusion types: Combining Tables 1 and 2 above and Figure 2 The experimental results are as follows: In fully occluded scenarios, the traditional method achieved an average recall of only 14.95% and a standard deviation as high as 2.98%, indicating that when the missing region is completely covered by the upper structure, the traditional strategy of relying solely on the circular buffer zone at the endpoints of structural units for local elevation determination is almost ineffective, and a large number of completely covered structures cannot be detected. In contrast, the method in this embodiment achieved an average recall of 96.79% and a standard deviation of only 0.96%, representing a recall improvement of approximately 6.5 times with minimal data fluctuation. In partially occluded scenarios, the method in this embodiment achieved an average recall of 94.24%, significantly better than the 42.68% of the traditional method. In structural fracture occlusion scenarios, the method in this embodiment achieved an average recall of 96.73%, also far exceeding the 36.97% of the traditional method.

[0099] The comparative experimental data above fully demonstrate that this embodiment, by constructing a topology map of the overpass structure and introducing the spatial location, semantic category, and topological connection relationship of each structural unit node as global context information into the occlusion determination process, can accurately identify occlusion occurring at any location of the structural unit. This fundamentally overcomes the inherent limitation of traditional local endpoint detection methods, which can only detect occlusion near endpoints, achieving a leapfrog improvement from local detection to global inference. Furthermore, the standard deviation of the method in this embodiment under all occlusion types is significantly smaller than that of traditional methods, exhibiting good detection stability and robustness. It does not suffer drastic performance degradation due to fluctuations in point cloud density or differences in occlusion morphology, thus providing an accurate and reliable prerequisite for subsequent conditional generative point cloud completion based on neighborhood constraints.

[0100] Please see Figure 3 The present invention also provides a system for constructing a three-dimensional model of an overpass, which is used to implement the above-mentioned method for constructing a three-dimensional model of an overpass, including: The point cloud segmentation and clustering module is used to perform semantic segmentation and instance segmentation on the original point cloud data, resulting in multiple structural unit point clouds with semantic category labels and instance labels; The topology graph construction module is used to treat each structural unit point cloud as a structural unit node based on the instance label, establish the topological connection relationship between each structural unit node based on the spatial position of each structural unit node, and construct the topology graph of the overpass structure. The missing occlusion determination module is used to determine the standard reference geometric information corresponding to each structural unit node based on the overpass structure topology map, compare each structural unit point cloud with the corresponding standard reference geometric information, identify the regions in which each structural unit point cloud is missing relative to the corresponding standard reference geometric information, and determine the occluded structural unit node and its occlusion type based on the spatial position of the region in the corresponding structural unit point cloud and the topological connection relationship of the corresponding structural unit node in the overpass structure topology map. The neighborhood constraint completion module is used to take the incomplete structural unit point cloud of the occluded structural unit node as input according to the occlusion type, and take the spatial position and geometric shape of each structural unit node adjacent to the occluded structural unit node in the overpass structural topology map as constraints. It is then input into the pre-trained conditional generative point cloud completion model for inference to generate the complete structural unit point cloud of the occluded structural unit node. The template matching and assembly module is used to match the complete structural unit point cloud with templates in the preset parametric structural template library, select the template with the highest matching degree and fit the corresponding geometric control parameters to generate parametric 3D models of each structural unit node; and assemble each parametric 3D model according to the corresponding spatial position according to the topological connection relationship to form a complete 3D semantic model of the overpass.

[0101] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0102] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0103] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0104] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for constructing a three-dimensional model of an overpass, characterized in that, The specific steps include: Semantic segmentation and instance segmentation are performed on the original point cloud data to obtain multiple structural unit point clouds with semantic category labels and instance labels; Based on the instance labels, each structural unit point cloud is treated as a structural unit node. The topological connection relationship between structural unit nodes is established according to the spatial location of each structural unit node, and the topological map of the overpass structure is constructed. Based on the topology diagram of the overpass structure, the standard reference geometric information corresponding to each structural unit node is determined. The point cloud of each structural unit is compared with the corresponding standard reference geometric information to identify the regions in which the point cloud of each structural unit is missing relative to the corresponding standard reference geometric information. Based on the spatial position of the region in the corresponding structural unit point cloud and the topological connection relationship of the corresponding structural unit node in the overpass structure topology diagram, the occluded structural unit node and its occlusion type are determined. Based on the occlusion type, the incomplete structural unit point cloud of the occluded structural unit node is taken as input, and the spatial position and geometric shape of each structural unit node adjacent to the occluded structural unit node in the overpass structural topology map are used as constraints. The inference is performed in the pre-trained conditional generative point cloud completion model to generate the complete structural unit point cloud of the occluded structural unit node. The complete structural unit point cloud is matched with templates in the preset parametric structural template library. The template with the highest matching degree is selected and the corresponding geometric control parameters are fitted to generate parametric 3D models of each structural unit node. Then, according to the topological connection relationship, each parametric 3D model is assembled according to its corresponding spatial position to form a complete 3D semantic model of the overpass.

2. The method for constructing a three-dimensional model of an overpass according to claim 1, characterized in that, The raw point cloud data is airborne lidar point cloud data; the semantic category labels include bridge deck category, ramp category, pier category, and guardrail category; The semantic segmentation and instance segmentation are achieved through a pre-trained 3D point cloud segmentation network. The 3D point cloud segmentation network extracts and classifies the semantic features of the original point cloud data point by point, and performs instance clustering based on the classification results and the spatial connectivity between point clouds, so as to distinguish the geometric regions of point clouds that are not connected under the same semantic category into different instances. The semantic segmentation and instance segmentation specifically include: first, performing point-by-point semantic classification on the original point cloud data, and assigning a corresponding semantic category label to each point cloud data point; For point cloud data that has been assigned semantic category labels, density-based connectivity clustering is performed based on the spatial location of each point cloud data point. Point cloud sets that belong to the same semantic category and are spatially continuous are divided into independent structural unit instances and assigned corresponding instance labels.

3. The method for constructing a three-dimensional model of an overpass according to claim 2, characterized in that, The topological connection relationship includes connected edges representing structural intersections and bifurcations, and intersecting edges representing overlapping upper and lower spatial layers; the structural intersection refers to the topological connection relationship in which multiple structural unit nodes converge to one structural unit node along the driving direction, and the bifurcation refers to the topological connection relationship in which one structural unit node branches into multiple structural unit nodes along the driving direction. The method for establishing the connecting edge is as follows: when the projections of two structural unit nodes in space overlap, and the structural unit point clouds corresponding to the two structural unit nodes in the overlapping area are connected or intersected in spatial position, a connecting edge is established between the two structural unit nodes. The method for establishing the cross edge is as follows: when the projections of two structural unit nodes in space overlap, and the point clouds of the structural unit nodes corresponding to the two structural unit nodes in the overlapping area have vertical spacing and do not intersect each other, a cross edge is established between the two structural unit nodes. Each structural unit node is associated with and stores the following attribute information: the spatial bounding box of the structural unit point cloud corresponding to the structural unit node, the structural orientation description information, the elevation description information, and the corresponding semantic category label; When the semantic category label is a bridge deck category or a ramp category, the structural orientation description information includes the centerline orientation or the trajectory line orientation; When the semantic category label is the bridge pier category, the structural orientation description information includes the vertical axis direction; When the semantic category label is the guardrail category, the structural orientation description information includes the horizontal trajectory orientation; The elevation description information includes elevation range or centerline elevation profile data.

4. The method for constructing a three-dimensional model of an overpass according to claim 3, characterized in that, The standard reference geometry information is determined based on the semantic category labels of the structural unit nodes: When the semantic category label of a structural unit node is a bridge deck category or a ramp category, the standard reference geometric information includes the reference centerline orientation, reference width variation range, and reference elevation variation range corresponding to the structural unit. When the semantic category label is a bridge pier, the standard reference geometric information includes the reference vertical axial orientation and reference cross-sectional dimension range corresponding to the structural unit; When the semantic category label is the guardrail category, the standard reference geometry information includes the reference horizontal trajectory and reference height profile range corresponding to the structural unit; For each point cloud data point in the point cloud of each structural unit, calculate the degree of spatial deviation between its spatial position and the standard position of the corresponding position in the standard reference geometric information; The continuous region formed by point cloud data points whose spatial deviation exceeds the preset tolerance range is identified as a missing region.

5. The method for constructing a three-dimensional model of an overpass according to claim 4, characterized in that, The types of occlusion include: When the spatial location of the missing region in the corresponding structural unit point cloud completely overlaps with the spatial projection area of ​​any structural unit node connected by a cross edge in the overpass structural topology map, it is determined to be a full occlusion type. When the missing region and the spatial projection area of ​​at least one structural unit node connected by intersecting edges in the overpass structure topology map only partially overlap, and there is no complete overlap with the spatial projection area of ​​any structural unit node, it is determined to be a partial occlusion type; the spatial projection area is the vertical projection coverage of the structural unit point cloud corresponding to each structural unit node on the horizontal plane, and its determination is based on the projection range of the spatial bounding box stored by each structural unit node on the horizontal plane. When the continuity of the connected edges of an obscured structural unit node in the overpass structural topology diagram is interrupted by a missing region, it is determined to be a structural break obscuration type. The determination method for the interruption of the continuity of the connected edges by a missing region is as follows: when the missing region divides the corresponding structural unit point cloud into at least two spatially unconnected point cloud segments along the driving direction, and the upstream and downstream structural unit nodes connected by the obscured structural unit node in the overpass structural topology diagram through connected edges are located exactly within these two point cloud segments, it is determined that the continuity of the connected edges is interrupted by a missing region. For full occlusion, the missing region spans the entire width of the corresponding structural unit point cloud; for partial occlusion, the missing region covers only a portion of the width or length of the corresponding structural unit point cloud; for structural break occlusion, the occluded structural unit node is connected to the upstream and downstream structural unit nodes respectively in the overpass structural topology diagram through connecting edges; the missing region divides the structural unit point cloud corresponding to the occluded structural unit node into at least two spatially unconnected point cloud segments.

6. The method for constructing a three-dimensional model of an overpass according to claim 5, characterized in that, The spatial location includes the spatial coordinates of each adjacent structural unit node, and the geometric shape includes the structural unit width, structural unit orientation vector, and corresponding semantic category label of each adjacent structural unit node. The conditional generative point cloud completion model is a deep learning network model built on a conditional generative adversarial network architecture or a conditional diffusion model architecture. It employs differentiated completion strategies based on different occlusion types, specifically including: When the occlusion type is full occlusion, the conditional generative point cloud completion model uses the spatial position and geometric shape of adjacent structural unit nodes as the only morphological reference to generate the complete structural unit point cloud of the occluded structural unit node. When the occlusion type is partial occlusion, the conditional generative point cloud completion model uses the geometry of the missing part in the incomplete structural unit point cloud of the occluded structural unit node as a reference, and combines the spatial position and geometry of the adjacent structural unit node to perform local completion and generate a complete structural unit point cloud. When the occlusion type is structural fracture occlusion, the conditional generative point cloud completion model takes the point cloud segments on both sides of the fracture of the occluded structural unit node as the reference, and the spatial position and geometry of the adjacent structural unit node as the guide to generate the intermediate transition segment point cloud connecting the two side point cloud segments.

7. The method for constructing a three-dimensional model of an overpass according to claim 6, characterized in that, The pre-training method for the conditional generative point cloud completion model is as follows: Multiple complete point cloud samples of interchange structural units are obtained. Each complete point cloud sample of interchange structural units is processed to simulate missing regions, generating incomplete point cloud samples of structural units with simulated missing regions. The point cloud samples of interchange structural units refer to a pre-constructed set of point cloud data of various types of structural units that are complete in geometric shape, including complete instance point clouds of bridge deck, ramps, piers and guardrails, and each sample has corresponding semantic category labels and instance labels. Using incomplete structural unit point cloud samples as input, corresponding complete overpass structural unit point cloud samples as supervision labels, and the spatial position and geometric shape of adjacent structural unit nodes in the simulated missing region in the overpass structural topology map as constraints, the conditional generative point cloud completion model is iteratively trained until the completion result output by the model meets the preset accuracy requirements.

8. The method for constructing a three-dimensional model of an overpass according to claim 7, characterized in that, After generating the complete structural unit point cloud of the occluded structural unit node, the following verification steps are also included: The generated complete structural unit point cloud is compared with the structural unit point cloud of the adjacent structural unit node at the spatial connection point. If the comparison result indicates that the geometric shape at the boundary is continuous and there is no spatial overlap, the completion result is confirmed to pass the verification. The geometric continuity refers to the fact that the average distance between the corresponding nearest neighbor pairs in the point clouds on both sides of the boundary is less than a first preset threshold, and the included angle of the tangential direction at the corresponding point is less than a second preset threshold; the spatial interlacing refers to the fact that the point clouds on both sides of the boundary penetrate each other, and the volume or number of point clouds in the penetrated area exceeds a third preset threshold. If the comparison results indicate that there are geometric discontinuities or spatial overlaps at the boundary, the boundary geometry of the adjacent structural unit nodes will be used as a mandatory constraint to correct the boundary of the completion result until the completion result passes the verification.

9. The method for constructing a three-dimensional model of an overpass according to claim 6, characterized in that, The preset parametric structure template library contains parametric geometric templates corresponding to various structural unit types, including: straight bridge deck templates, curved ramp templates, bifurcation structure templates, and intersection structure templates; each parametric geometric template corresponds to a structural unit type, and the geometric shape of the structural unit type is defined by a set of adjustable geometric control parameters. The specific matching method is as follows: calculate the geometric feature similarity between the complete structural unit point cloud and the calibration geometric shape defined by each parameterized geometric template, and select the parameterized geometric template with the highest geometric feature similarity as the matching result; the geometric feature similarity includes the overall orientation similarity, width distribution similarity and elevation change similarity of the structural unit, and the similarity of the three dimensions is weighted and combined to obtain the final similarity score; Based on the geometric control parameters defined by the parameterized geometric template selected according to the matching results, the corresponding geometric feature values ​​are extracted from the point cloud of the complete structural unit, and the extracted geometric feature values ​​are used as the geometric control parameters of the structural unit node; the geometric feature values ​​include the centerline curvature radius, cross-sectional width, elevation change and end face tilt angle of the structural unit; The specific assembly method is as follows: Based on the adjacent structural unit nodes connected by each connected edge in the overpass structure topology diagram, determine the common connection end face between adjacent parametric 3D models; using the spatial position of each common connection end face as the alignment reference, perform spatial transformation and positioning on each parametric 3D model to achieve the alignment and splicing of adjacent end faces; After forming a complete three-dimensional semantic model of the overpass, the following verification steps are also included: traversing each intersection edge in the topology diagram of the overpass structure, checking whether there is spatial penetration or interference in the vertical direction of each structural unit model in the complete three-dimensional semantic model of the overpass; if there is spatial penetration or interference, the bottom surface of the upper structural unit model connected by the intersection edge is used as a constraint to avoid and correct the top surface of the lower structural unit model.

10. A system for constructing a three-dimensional model of an overpass, characterized in that, The system for constructing a three-dimensional model of an overpass is used to implement the method for constructing a three-dimensional model of an overpass as described in any one of claims 1-9, comprising: The point cloud segmentation and clustering module is used to perform semantic segmentation and instance segmentation on the original point cloud data, resulting in multiple structural unit point clouds with semantic category labels and instance labels; The topology graph construction module is used to treat each structural unit point cloud as a structural unit node based on the instance label, establish the topological connection relationship between each structural unit node based on the spatial position of each structural unit node, and construct the topology graph of the overpass structure. The missing occlusion determination module is used to determine the standard reference geometric information corresponding to each structural unit node based on the overpass structure topology map, compare each structural unit point cloud with the corresponding standard reference geometric information, identify the regions in which each structural unit point cloud is missing relative to the corresponding standard reference geometric information, and determine the occluded structural unit node and its occlusion type based on the spatial position of the region in the corresponding structural unit point cloud and the topological connection relationship of the corresponding structural unit node in the overpass structure topology map. The neighborhood constraint completion module is used to take the incomplete structural unit point cloud of the occluded structural unit node as input according to the occlusion type, and take the spatial position and geometric shape of each structural unit node adjacent to the occluded structural unit node in the overpass structural topology map as constraints. It is then input into the pre-trained conditional generative point cloud completion model for inference to generate the complete structural unit point cloud of the occluded structural unit node. The template matching and assembly module is used to match the complete structural unit point cloud with templates in the preset parametric structural template library, select the template with the highest matching degree and fit the corresponding geometric control parameters to generate parametric 3D models of each structural unit node; and assemble each parametric 3D model according to the corresponding spatial position according to the topological connection relationship to form a complete 3D semantic model of the overpass.

Images