Railway scene-oriented live-action three-dimensional lightweight method

Abstract

The invention discloses a live-action three-dimensional lightweight method for a railway scene. The method comprises the following steps: acquiring live-action three-dimensional image data in a preset area containing a railway main body by utilizing unmanned oblique photography; performing pretreatment; inputting the DOM data into the trained deep learning classification model, and carrying out ground feature category segmentation and semantic processing to obtain monotonized model data of different categories and different blocks; performing partition compression on each type of monomer model data; constructing a dynamic LOD system based on an octree according to different category partitions; and according to the established dynamic LOD system, generating a unique identifier for each block by adopting a spatial index tree, establishing an R tree spatial index, performing fusion processing on adjacent block joints by adopting a multi-resolution model, and performing merging optimization on block data. According to the method, high-efficiency lightweight of the railway live-action three-dimensional model is realized, the data processing and application efficiency is improved, and powerful support is provided for digital transformation of railway engineering.

Description

Technical Field

[0001] The present invention relates to the technical field of geographic information and railway data processing, and particularly to a real-scene three-dimensional lightweight method for railway scenarios. Background Art

[0002] With the rapid development of geographic information technology and railway construction, the application of real-scene three-dimensional models in railway engineering has become increasingly widespread. High-precision real-scene three-dimensional data of railway areas can be obtained through technologies such as unmanned oblique photography, providing rich information support for railway planning, design, construction, and operation and maintenance. However, real-scene three-dimensional data is characterized by large data volume and high complexity, posing huge challenges to data storage, transmission, and processing. In railway scenarios, traditional real-scene three-dimensional model processing methods have some deficiencies. There may be redundant data such as floating objects in the model, affecting the accuracy and practicality of the model. The classification of ground objects is not fine enough, which is not conducive to targeted processing of different types of ground objects; the model compression lacks an effective weight grading strategy, easily resulting in the loss of accuracy and details of important ground objects, while the compression of some unimportant ground objects is not sufficient, causing data redundancy. During the process of merging block data, the lack of efficient spatial indexing and data fusion methods leads to low data retrieval and processing efficiency. The processing of the top-level structure is not perfect, and problems such as redundant vertex indexing affect the overall performance of the model. Therefore, a real-scene three-dimensional lightweight method for railway scenarios is needed to solve the above problems. Summary of the Invention

[0003] Therefore, the purpose of the present invention is to provide a real-scene three-dimensional lightweight method for railway scenarios, which can effectively eliminate redundant data, classify ground objects finely, rationally formulate compression weights, optimize the merging of block data and the processing of the top-level structure, so as to achieve the efficient lightweight of railway real-scene three-dimensional models, improve the efficiency of data processing and application, and provide strong support for the digital transformation of railway engineering.

[0004] To achieve the above purpose, a real-scene three-dimensional lightweight method for railway scenarios proposed by the present invention includes the following steps: S1. Use unmanned oblique photography to obtain real-scene three-dimensional image data within a preset area of the railway main body; S2. Preprocess the obtained real-scene three-dimensional data image data to eliminate noise data; S3. Convert the format of the preprocessed three-dimensional data image data to obtain DOM data, input the DOM data into a trained deep learning classification model, perform ground object category segmentation and semantic processing, and obtain monomeric model data of different categories and different blocks; S4. Perform partition compression on each type of monomer model data according to the dynamically obtained weights according to the classification; S5. Construct a dynamic LOD system based on octree by partitioning into different categories; S6. Based on the established dynamic LOD system, use a spatial index tree to generate a unique identifier for each block, establish an R-tree spatial index, use a multi-resolution model to fuse and process the seams of adjacent blocks, and merge and optimize the block data.

[0005] Further preferably, in S2, the preprocessing of the acquired real-scene three-dimensional data image data to remove noise data includes using an outlier detection method to remove airborne suspended matter. Among them, the outlier detection method includes: Calculate the distance from the centroid p of any real-scene three-dimensional model block to its K-nearest neighbor, denoted as the K-distance; Calculate the reachable distance R_D based on the K-distance; Calculate the local reachability density LRD ( ); Calculate the local outlier factor according to the local reachability density LRD ; Make a judgment according to the set threshold. When the local outlier factor is greater than the set threshold, the current real-scene module is determined as the target to be deleted; otherwise, the real-scene module is retained; Among them, is the centroid of any real-scene three-dimensional model block; is the centroid of the real-scene three-dimensional model block adjacent to P i , k is the total number of centroids of the real-scene three-dimensional model blocks adjacent to , i is the serial number of any centroid, and j is the serial number of the centroid of the real-scene three-dimensional model block adjacent to .

[0006] Further preferably, in S3, the deep learning classification model is trained based on labeled samples; Edit the vectorized classification results through QGIS, edit the ID and category of each object, and obtain the vector data of the object ID, category type, and regional ring coordinate string polygon of different blocks; Use the partitioned vector data to perform block cutting and monomerization processing on the real-scene three-dimensional model to obtain the monomerized model data of different categories and different blocks.

[0007] Further preferably, in S4, compress each monomer model data according to the weights dynamically obtained by classification, including performing geometric compression and texture compression on each category according to different types; the geometric compression uses the vertex folding method to calculate whether the original cost of the new vertex after folding meets the error threshold , and the error threshold is calculated according to the weights dynamically obtained by classification: L; L is the length of the diagonal of the bounding box; W final is the final weight; Calculate the final weight W of each type of object according to the following formula final ; where, W static is the static weight, F density is the patch density factor, F complexity is the geometric complexity factor; The patch density factor F density is calculated using the following formula: where, N class is the number of patches of the current category, N total is the total number of patches, and the threshold is set to 30%.

[0008] The geometric complexity factor F complexity is calculated using the following formula: is the standard deviation of the local Gaussian curvature, is the mean value, and the calculation window is a radius r = 0.1 × the length L of the diagonal of the bounding box.

[0009] Further preferably, in S5, construct an octree-based dynamic LOD system according to different category partitions; among them, the LOD level is calculated according to the following formula: where, L is the side length of the bounding box, and D is the minimum cell size.

[0010] Further preferably, in S6, the spatial index tree is used to generate a unique identifier for each block and establish an R-tree spatial index, including the following steps: Adopt the improved linear octree coding, divide the management unit according to a 100m × 100m grid, and establish a four-dimensional index (x, y, z, type) to realize spatial-attribute joint retrieval; Generate a unique UUID identifier for each block and establish an R-tree spatial index; The rules of the improved linear octree coding are as follows: where, the octant number OctantID of each level s [0, 7], is the total number of LOD levels, and s is the block number of the spatial index tree.

[0011] Further preferably, in S6, the multi-resolution model fusion for processing adjacent block seams includes: Judging whether the Euclidean distance between any two adjacent spatial block models exceeds a limit value; when it exceeds the limit value, insert transitional triangular patches to adjust the Euclidean distance between the two adjacent spatial block models to meet the requirement of being less than the limit value; Unify the coordinate system using the vertex remapping algorithm shown below, and convert the vertex coordinates of each block model to the global coordinate system: where R is the block rotation matrix, and O block is the origin of the block local coordinate system, is the translation amount of the block in the global coordinate system; is the original coordinate; Calculate the intersection of the bounding boxes of all blocks to determine the global origin ; N is the total number of blocks; Perform coordinate transformation, and check whether the coordinate difference after quantization of the adjacent block boundary fixed points meets the coordinate difference ； If not, recheck other adjacent blocks; When it meets the requirement, merge redundant vertex data, and adopt an adaptive texture atlas merging strategy to generate a composite texture atlas.

[0012] Further preferably, the merging and optimization of the block data includes: Construct a vertex-triangle adjacency list in the following form: calculate the vertex score; ; represents any vertex, and l represents the vertex number of the vertex-triangle; always select the vertex with the highest current score to be added to the cache, and update the adjacency list; Generate an optimized index buffer and output a new index sequence I optimized to ensure that adjacent triangles share vertices; Construct a hierarchical detail transition LOD parameter matrix to ensure visual continuity.

[0013] The real-scene three-dimensional lightweight method for railway scenes disclosed in this application has at least the following advantages compared with the prior art: In terms of data quality, the accuracy and practicality of the model are improved by automatically removing floating objects and performing fine-grained ground object classification, enabling the real-scene three-dimensional model of the railway scene to more accurately reflect the actual situation. In terms of data processing efficiency, the zoning weight formulation and multi-resolution hybrid compression technology are adopted to perform targeted compression on different ground object categories, significantly reducing the data volume while maximizing the preservation of the accuracy and details of important ground objects such as the main railway project. During the process of block data merging and top-level reset processing, the constructed efficient spatial index and the adopted fusion and optimization algorithms achieve fast data retrieval, accurate fusion, and unified management, significantly improving the efficiency of data storage, transmission, and processing. This method effectively solves the problems of large volume and difficult processing of railway real-scene three-dimensional data, providing strong support for the digital construction of railway projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 FIG. is a flowchart framework diagram of a real-scene three-dimensional lightweight method for railway scenes proposed by the present invention.

[0015] Figure 2 FIG. is a detailed flowchart diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] As Figure 1 shown, an embodiment of the present invention provides a real-scene three-dimensional lightweight method for railway scenes, including the following steps: S1. Use unmanned oblique photography to obtain real-scene three-dimensional image data within a preset area of the main railway body, including: flying at a designed altitude of 100 - 150m, with a ground resolution (GSD) ≤ 3cm, a forward overlap of 80%, and a side overlap of 70% to ensure the three-dimensional reconstruction accuracy. Covering the main railway project and a 150m range on both sides, a multi-view image photo is obtained using the grid flight path + key area surrounding mode, and the real-time position and attitude data are recorded through the POS system. Through three-dimensional reconstruction software, multi-view image matching and three-dimensional reconstruction are performed, and real-scene three-dimensional model data in the OSGB format of the railway area is output. The data includes the main railway project and the data of the main railway facilities and the external environment within a 150m range on one side of the railway. This process corresponds to Figure 2 the unmanned aerial vehicle oblique photography in the flowchart shown.

[0018] S2. Preprocess the obtained real-scene three-dimensional data image data to remove noise data, including using an outlier detection point cloud analysis algorithm to remove aerial suspended objects and manually review key areas to ensure no omission within the railway boundary. That is, Figure 2 the reconstruction and preprocessing process of the model in FIG., including the removal of floating objects and the OSGB after the removal of floating objects; Among them, the specific detection process of the outlier detection point cloud analysis algorithm is as follows: (1) Calculate the K-distance and reachability distance: The K-distance \(K_D\) is the distance from point \(p\) i to its \(k\)-th nearest neighbor; The reachability distance \(R_D(p i , p j ) = max(k_D, ||p i - p j ||); ||p i - p j || represents the Euclidean distance between two points \(p i \) and \(p j ; (2) Local reachability density (LRD) (3) Local outlier factor (LOF): If LOF(p i ) > 1.5, then the real scene model block represented by \(p i \) is the target to be deleted; Among them, is the centroid of any real scene three-dimensional model block; is the centroid of the real scene three-dimensional model block adjacent to \(P i \), \(k\) is the total number of centroids of the real scene three-dimensional model blocks adjacent to \); \(i\) is the serial number of any centroid, and \(j\) is the serial number of the centroid of the real scene three-dimensional model block adjacent to \).

[0019] S3. Perform format conversion on the preprocessed three-dimensional data image data to obtain DOM data, input the DOM data into the trained deep learning classification model, and perform ground object category segmentation and semantic processing to obtain the monomerized model data of different categories and different blocks. That is Figure 2 the classification of ground objects and the process of model monomerization in, including converting to DOM format - ground object classification - obtaining category vectors; using the OSGB obtained in S2 to perform model segmentation - obtaining the object monomerized OSGB model, where the monomerized OSGB model includes the main body of the railway, ancillary facilities, water bodies, vegetation, and artificial structures.

[0020] Specifically, it includes: inputting DOM data into the trained U-Net deep learning classification model, using the model to perform land cover type segmentation and semanticization on DOM, editing the vectorized classification results through QGIS, modifying incorrect classification boundaries and incorrect class information, editing the ID and class (type) of each object, and obtaining vector data of object IDs, class types, and regional ring coordinate strings (polygons) of different blocks. Using the partitioned vector data, the real-scene 3D model is segmented, cut, and monomerized to obtain monomerized model data of different classes and different blocks.

[0021] S4. According to the weights dynamically obtained by classification, partitioned compression is performed on the monomer model data of each type. Corresponding to the formulation of the compression partition weights and the algorithm strategy in the appendix Figure 2 including: weight formulation - generating weight files (including: main project, auxiliary facilities, artificial structures, water bodies and vegetation). After each type of weight file is set, patch compression (i.e., geometric compression) and texture compression are performed; the specific compression methods are as follows: It should be noted that when performing compression on the monomer model data of each type, it includes two aspects of geometric compression and texture compression for each category according to different types: For geometric compression, the vertex folding method is adopted: First, vertex folding is performed in the 3D network; vertex folding is a key operation in the 3D network, aiming to reduce the number of vertices of the model and retain geometric features as much as possible, improving rendering efficiency and reducing storage costs; the specific process includes the following: Select vertex pairs: Determine the vertices to be folded (V i , V j ) and their target position, the new vertex V new .

[0022] Calculate the folding cost: Use the quadratic error matrix (QEM) to measure the impact of folding on the geometric error of the model.

[0023] Quadratic error matrix (QEM) calculation: For each vertex v, calculate the plane equation ax + by + cz + d = 0 of its associated face, and the error matrix Q v is: The error cost of the new vertex V new after folding is:

[0024] where Q i , Q j are the quadratic error matrices of Vi and Vj respectively, and the smaller the value, the smaller the error introduced by folding.

[0025] Perform folding: Fold V i to the target position, update the vertex, edge, and face structure of the mesh, and remove the faces and edges degenerated due to folding.

[0026] Iterative simplification: Repeat the above steps until the mesh simplification meets the error threshold . The error threshold is calculated based on the weights dynamically obtained according to the classification L; L is the length of the diagonal of the bounding box; W final is the final weight.

[0027] The weights dynamically obtained according to the classification include the weights of each type of object set as follows: The final weight W of each type of object final consists of the static weight W static , the patch density factor F density and the geometric complexity factor F complexity for dynamic adjustment: The patch density factor is calculated using the following formula: where N class is the number of patches of the current category, N total is the total number of patches, and the threshold is set to 30%.

[0028] The geometric complexity factor is calculated using the following formula: is the standard deviation of the local Gaussian curvature, is the mean value, and the calculation window is a radius r = 0.1 × the length L of the diagonal of the bounding box.

[0029] The static weights of each type of object are in accordance with the following convention: ① For the main railway project category Set = 0.8, and the final weight W final is determined according to the patch density factor and the geometric complexity factor to ensure that its accuracy and details are retained to the greatest extent during the model compression process. A relatively conservative edge folding triangular patch compression algorithm is adopted, and the error threshold T is set to ensure that the geometric shape and key features of the main project are not damaged. The specific calculation method is as follows: The folding constraint condition is: When the Gaussian curvature K of the vertex v satisfies: and satisfies: L, fold its associated edge.

[0030] Among them, is the global curvature mean and standard deviation, and L is the length of the diagonal of the bounding box.

[0031] For texture compression, the mildly lossy DXT5 compression algorithm is adopted to maintain the clarity and authenticity of the texture. The threshold standard T for DXT compression quality DXT = 40 * W final , and the compression of the data should satisfy PSNR > T DXT .

[0032] ② Auxiliary facilities Set = 0.5, and the final weight W final is determined according to the patch density factor and the geometric complexity factor. In triangular patch compression, the adaptive triangular mesh compression algorithm is used to set the error threshold T to ensure the accuracy of key components such as signal and communication equipment. The specific calculation method: the error of triangular mesh adaption When L, perform the merging of patches.

[0033] Texture compression adopts the medium - degree DXT3 lossy compression algorithm to reduce the texture data volume to a certain extent while maintaining a good visual effect. The threshold standard T for DXT compression quality DXT = 40 * W final , and the compression of the data should satisfy PSNR > T DXT .

[0034] ③ Man - made structures Different weights are assigned according to the importance and function of specific facilities. Set = 0.3, and the final weight W final is determined according to the patch density factor and the geometric complexity factor. The region - growing clustering algorithm is adopted, and the clustering radius r cluster is calculated as: V NN is the nearest point to point V i . When the normal vector angle θ < 15°, and the change rate δ of the patch area after merging area < 20%, execute the clustering algorithm. Texture compression adopts the DXT3 compression strategy to further reduce the model data volume. The threshold standard T for DXT compression quality DXT = 40 * W final , and the compression of the data should satisfy PSNR > T DXT .

[0035] ④ Natural features such as water bodies and vegetation Different weights are assigned according to the importance and function of specific facilities. Set = 0.1, and the final weight Wfinal Determined according to the patch density factor and the geometric complexity factor. A larger simplification ratio can be adopted for compression, and the region growing algorithm is used in combination with a larger tolerance to merge similar patches. V NN Is the nearest point to point V i , when the angle θ between the normal vectors is < 30°, and the change rate δ of the patch area after merging area < 40%, the clustering algorithm is executed.

[0036] Texture compression can adopt the highly lossy DXT1 format compression algorithm. By adjusting the compression quality parameter, the texture data volume can be significantly reduced without significantly affecting the overall visual effect. The threshold standard T for DXT compression quality DXT = 40 * W final , and the compression of the data should satisfy PSNR > T DXT .

[0037] DXT compression quality evaluation: , where MSS is the root mean square error of texture compression.

[0038] S5. Construct a dynamic LOD system based on octree by partitioning according to different categories; see Figure 2 for the differential LOD partition compression processing process in it. Among them, for the models of the main railway project category, the original LOD levels (0 - 25 levels) are retained, for the affiliated facilities category, 0 - 22 levels are retained, for the artificial structures, 0 - 20 levels are retained, and for the natural features category, only 0 - 15 levels are retained. The calculation method of the specific LOD levels is as follows: where L is the side length of the bounding box, D is the minimum cell size, and the LOD level retention rules are as shown in Table 1 below: Table 1 LOD level retention rules table .

[0039] S6. Based on the established dynamic LOD system, use a spatial index tree to generate a unique identifier for each block, establish an R-tree spatial index, and use a multi-resolution model to fuse the seams of adjacent blocks to merge and optimize the block data. That is, Figure 2 the data merging and optimization in it include constructing a spatial index tree structure, multi-resolution model fusion, and data integration optimization, etc.; Adopt the improved linear octree coding (LOCT), divide the management unit according to a 100m × 100m grid, establish a four-dimensional index (x, y, z, type) to realize spatial-attribute joint retrieval, generate a unique UUID identifier for each block and establish an R-tree spatial index. The specific coding rules are as follows: Among them, each hierarchical octant is numbered OctantID s [0, 7], is the total number of LOD levels, and s is the block number of the spatial index tree.

[0040] Judge whether the Euclidean distance between any two adjacent spatial block models exceeds the limit value; when it exceeds the limit value, insert transitional triangular patches to adjust the Euclidean distance between the two adjacent spatial block models to meet the requirement of being less than the limit value; Use the TopoJSON algorithm for topology-preserving boundary matching to process the seams between adjacent blocks. The specific calculation method is as follows: B i , B j are two adjacent spatial block models (Blocks). Block 's set of boundary fixed points.

[0041] : Any fixed point on the boundary of block with coordinates (x v , y v , z v ).

[0042] : The theoretical matching point corresponding to v on the boundary of block with coordinates (x v , y v , z v ).

[0043] The maximum allowable error for fixed point alignment is set to 0.001m.

[0044] For any two adjacent blocks B i and B j , it is necessary to satisfy: That is, for each apex V on the boundary of block B i , there must exist a vertex V' on the boundary of the adjacent block B j , and the Euclidean distance between the two does not exceed the threshold .

[0045] If there exists a vertex pair (v, v') that does not meet the above conditions, transitional triangular patches need to be inserted. The specific steps are as follows: Define the offset vector , if , then correction is required.

[0046] Adjust the coordinates of v and v' by the least squares method to minimize the offset: Subject to

[0047] where v original , v' original are the original fixed-point coordinates.

[0048] 3) Apply vertex remapping technology to unify the coordinate system and convert the vertex coordinates of each sub-block model to the global coordinate system: where, is the original coordinate, R is the sub-block rotation matrix, O block is the origin of the sub-block local coordinate system, is the translation amount of the sub-block in the global coordinate system.

[0049] Calculate the intersection of the bounding boxes of all sub-blocks to determine the global origin , and N is the total number of sub-blocks.

[0050] Perform coordinate transformation and check the coordinate difference of the quantized boundary fixed points of adjacent sub-blocks 。 If not satisfied, re-check other adjacent sub-blocks Use InstancedMesh technology to merge redundant vertex data and adopt an adaptive texture atlas merging strategy to generate a 2048×2048 composite texture atlas.

[0051] 4) Data integration and optimization, improve vertex cache optimization and hierarchical detail transition Implement vertex cache optimization (VBO) to reorganize geometric data. Construct a vertex-triangle adjacency list: represents any vertex, l represents the vertex number of the vertex-triangle; Calculate the vertex score: is the area of the triangle; Age(v) is the number of unused steps since the vertex was last accessed, and k is the cache capacity. Always select the vertex with the highest current score to join the cache and update the adjacency list. Generate an optimized index buffer and output a new index sequence I optimized , ensuring that adjacent triangles share vertices. Verify ACMR, requiring ACMR ≤ 1.5 after optimization.

[0052] Use instanced rendering technology to process repeated components, calculate the component hash, and calculate a unique identifier for each Mesh: ; represents the encryption level, is the vertex data, is the index data, is the material file ID.

[0053] Instantiation is started when the following conditions are met: ≥0.1 Construct a hierarchical level of detail transition (LOD Transition) parameter matrix to ensure visual continuity. Define the viewing distance threshold and set the switching distance for each LOD level according to the radius R of the model bounding sphere: LOD i is the i-th layer of LOD; is the maximum number of layers.

[0054] Define the transition interval [d start , d end : Calculation of the mixing factor α: is the smooth interpolation function, and d is the interpolation of the natural transition generated between ...

[0055] Obviously, the above embodiments are only examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A real - scene three - dimensional lightweight method for railway scenarios, characterized in that, The following steps are involved: S1. Using unmanned oblique photography to obtain real-scene three-dimensional image data within a preset area of the railway body; S2, pre-processing the acquired real-scene three-dimensional data image data to remove noise data; S3, converting the format of the pre-processed three-dimensional image data to obtain DOM data, inputting the DOM data into the trained deep learning classification model, performing ground object category segmentation and semantic processing, and obtaining individual model data of different categories and different blocks; S4. Partition and compress the data of each type of monomer model according to the weights dynamically obtained by classification; calculate the final weight W of each type of object according to the following formula final； ； Among them, W static is the static weight, F density is the patch density factor, F complexity is the geometric complexity factor; S5. Construct a dynamic LOD system based on octree according to different category partitions; S6. Based on the established dynamic LOD system, a spatial index tree is used to generate a unique identifier for each block, an R-tree spatial index is established, and a multi-resolution model is used to fuse the seams of adjacent blocks to merge and optimize the block data.

2. The real-scene three-dimensional lightweight method for railway scenarios according to claim 1, wherein In S2, the acquired real-scene three-dimensional data image data is preprocessed to remove noise data, including using an outlier detection method to remove suspended objects in the air, wherein the outlier detection method includes: Calculate the distance from the centroid p of any real-scene 3D model block to the Kth nearest neighbor, recorded as K-distance; Calculate the reachable distance R_D based on K-distance; Calculate the local reachability density LRD according to the reachable distance R_D( ); Calculate the local outlier factor according to the local reachability density (LRD). ; Judgment is made according to the set threshold. When the local outlier factor is greater than the set threshold, the current real scene module is determined as the target to be deleted; otherwise, the real scene module is retained; Among them, is the centroid of any real scene 3D model block; is the centroid of the real scene 3D model block adjacent to P i , k is the total number of centroids of the real scene 3D model blocks adjacent to ; i is the serial number of any centroid, and j is the serial number of the centroid of the real scene 3D model block adjacent to .

3. The real-scene three-dimensional lightweight method for railway scenarios according to claim 1, characterized in that In S3, the deep learning classification model is trained based on the labeled samples; Edit the vector classification results through QGIS, edit the ID and category of each object, and obtain the object ID, category type and regional ring coordinate string polygon vector data of different blocks; The partition vector data is used to cut the real-life 3D model into blocks and individualize it to obtain individual model data of different categories and blocks.

4. The real-scene three-dimensional lightweight method for railway scenarios according to claim 1, wherein In S4, each single model data is compressed according to the weights dynamically obtained by classification, including geometric compression and texture compression for each category according to different types; the geometric compression uses the vertex folding method to calculate whether the original cost of the new vertex after folding meets the error threshold , the error threshold is calculated according to the weights dynamically obtained by classification: L; L is the length of the diagonal of the bounding box; W final is the final weight; Patch density factor F density It is calculated using the following formula: Among them, N class is the number of patches for the current category, N total is the total number of patches, and the threshold is set to 30%; Geometric complexity factor F complexity It is calculated using the following formula: is the standard deviation of the local Gaussian curvature, and is the mean value.

5. The real-scene three-dimensional lightweight method for railway scenarios according to claim 1, characterized in that In S5, a dynamic LOD system based on octree is constructed according to different category partitions; the LOD level is calculated according to the following formula: Among them, L is the side length of the bounding box and D is the minimum unit size.

6. The real-scene three-dimensional lightweight method for railway scenarios according to claim 1, characterized in that In S6, the spatial index tree is used to generate a unique identifier for each block, and an R-tree spatial index is established, including the following steps: The improved linear octree coding is used to divide the management unit into 100m×100m grids, and a four-dimensional index (x, y, z, type) is established to realize space-attribute joint retrieval; Generate a unique UUID identifier for each block and create an R-tree spatial index; The improved linear octree encoding rules are as follows: Among them, the OctantID of each hierarchical octant s [0, 7], is the total number of LOD levels, and s is the block number of the spatial index tree.

7. The real-scene three-dimensional lightweight method for railway scenarios according to claim 1, characterized in that In S6, the multi-resolution model fusion is used to process the seams of adjacent blocks, including: Determine whether the Euclidean distance between any two adjacent spatial block models exceeds a limit value; if the limit value is exceeded, insert transition triangles to adjust the Euclidean distance between the two adjacent spatial block models to meet the requirement of being less than the limit value; The vertex remapping algorithm shown below is used to unify the coordinate system and transform the vertex coordinates of each block model into the global coordinate system: Among them, R is a block rotation matrix, and O block is the origin of the block local coordinate system, is the translation amount of the block in the global coordinate system; is the original coordinate; Calculate the bounding box intersection of all blocks and determine the global origin ; N is the total number of blocks; Perform coordinate transformation and check whether the coordinate differences after quantization of the fixed points at the boundaries of adjacent blocks satisfy the coordinate differences ； If not, recheck other adjacent blocks; When the conditions are met, the redundant vertex data are merged and an adaptive texture atlas merging strategy is adopted to generate a composite texture atlas.

8. The real-scene three-dimensional lightweight method for railway scenarios according to claim 1, characterized in that The merging and optimizing of the block data includes: Construct a vertex-triangle adjacency list of the following form: Calculate the vertex score; ; represents any vertex, and l represents the vertex number of the vertex-triangle; Always select the vertex with the highest current score to add to the cache and update the adjacency list; Generate an optimized index buffer and output a new index sequence I optimized , ensuring that adjacent triangles share vertices; Constructs a layer-of-detail transition LOD parameter matrix to ensure visual continuity.

Images