A multi-level detail automatic simplification method for oblique photography real scene three-dimensional model

Abstract

The application discloses a kind of oblique photography real scene three-dimensional model multi-level detail automatic simplification method, it is related to three-dimensional model simplification and computer graphics technical field, comprising: obtaining oblique photography original data and three-dimensional model basic information;Model pre-processing, self-adapting gauss filter denoising, improved RANSAC removes outlier, block compression texture and correct mapping coordinates, repair topological structure;Extract multi-scale feature;Simplification decision model is constructed, and the simplification rate and priority are determined in combination with observation distance, scene accuracy and hardware performance;Perform hierarchical simplification, vertex level improves edge folding, face sheet level self-adapting deletion, regional level grid reconstruction;Multi-dimensional quality assessment, then backtrack adjustment if not meet requirement;Output is stored in the simplification model according to LOD level, including transition information and simplification log.The application improves data quality by elaboration preprocessing, multi-dimensional feature extraction and intelligent decision enhance simplification pertinence, improve model application value.

Description

Technical Field

[0001] This invention relates to the fields of 3D model simplification and computer graphics technology, and in particular to a method for automatically simplifying multi-level details of oblique photogrammetry real-world 3D models. Background Technology

[0002] Oblique photogrammetry, by fusing multi-view images with laser point clouds, can construct high-fidelity realistic 3D models, widely used in smart cities, surveying and mapping, digital twins, and other fields. However, these models have extremely large data volumes; a single region model often contains millions or even hundreds of millions of vertices and faces. After overlaying high-resolution textures, the data volume can reach tens of gigabytes or even terabytes. Existing model simplification methods have significant shortcomings in the preprocessing stage. Traditional Gaussian filtering uses a fixed kernel size, resulting in uneven denoising effects for models with large differences in vertex density (such as densely built-up areas versus open terrain). Dense areas tend to blur details, while sparse areas still retain noise. Outlier removal relies on simple distance thresholds, making it difficult to distinguish between real outliers and effective features such as terrain protrusions, leading to chaotic topological structures in subsequent models. Texture compression often uses a uniform compression ratio without considering the matching relationship between face area and texture resolution. Small faces corresponding to high-resolution textures result in memory waste, while large faces with low-resolution textures exhibit significant stretching and distortion, affecting visual effects.

[0003] Existing simplification methods suffer from limitations in feature extraction and decision-making mechanisms that cannot adapt to the complexity of oblique photogrammetry models. Feature extraction is often limited to vertex-level normals and curvatures, neglecting key information such as patch visibility and regional semantics. For example, using the same simplification strategy for building facades and vegetation areas leads to oversimplification of key structures such as building edges and distortion of vegetation outlines. Simplification decisions rely solely on observation distance, failing to consider the accuracy requirements of the application scenario and differences in hardware performance. Building surveying scenarios require centimeter-level facade details, while macro-planning scenarios have lower requirements for terrain accuracy, but existing methods cannot dynamically adjust these requirements. Hardware adaptability is also insufficient; high-end equipment does not fully utilize its performance, resulting in conservative simplification, while low-end equipment experiences lag due to incomplete simplification. Furthermore, the lack of differentiated simplification thresholds for different semantic regions (such as buildings, roads, and water bodies) further reduces the rationality and practicality of model simplification.

[0004] The shortcomings in simplified operation and quality assessment further restrict the application effect of the model. The edge folding algorithm at the vertex level does not consider semantic importance, and building corner points are folded in the same way as ordinary wall vertices, resulting in deformation of the building structure; the face deletion does not check the continuity of the normal vectors of the remaining faces, which is prone to local concavity or convexity; the region level lacks a targeted reconstruction strategy, and flat terrain and complex buildings are treated with the same mesh processing method, making it difficult to balance efficiency and accuracy. The quality assessment dimension is too single, only calculating Hausdorff distance and ignoring key indicators such as volume deviation and texture matching degree. Although some simplified models meet the distance requirements, the volume deviation is too large or the texture is misaligned, which cannot meet the needs of practical applications; the LOD level transition lacks a smooth mechanism, and there are obvious jumps when switching, which affects the visual coherence of real-time rendering. Moreover, the simplification process is not logged, making it difficult to trace the root cause of the problem and hindering subsequent optimization iterations. Summary of the Invention

[0005] This invention proposes an automatic simplification method for multi-level details of oblique photogrammetry real-world 3D models to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention employs the following technical solution: a method for automatically simplifying multi-level details of a 3D oblique photogrammetry real-world model, comprising:

[0007] Acquire raw oblique photogrammetry data and basic information of the 3D model; the raw data includes multi-view oblique images, laser point clouds, POS data and camera intrinsic parameters, and the basic information of the 3D model includes vertex coordinates, triangular facet indexes, texture coordinate mapping relationships, initial LOD level division and model bounding box parameters; store the two types of data through spatial coordinate association, establish a globally unique identifier mapping table, and record the data acquisition time, equipment model and accuracy level.

[0008] The 3D model is preprocessed; adaptive Gaussian filtering is used to denoise the vertices, the RANSAC algorithm is improved to remove outliers of the laser point cloud, the texture image is compressed in blocks and the mapping coordinates are corrected, the model topology is repaired, and duplicate vertices, non-manifold edges and degenerate faces are removed.

[0009] Extract multi-scale features from the 3D model; calculate parameters at the vertex level, calculate indices at the face level, and divide regions using a feature clustering algorithm at the region level.

[0010] Construct a simplified decision-making model; comprehensively consider observation distance, application scenario accuracy, hardware performance, and semantic features, set regional differentiation thresholds and minimum simplification rates, and output the simplification priority of each region through a multi-factor weighted algorithm;

[0011] Perform hierarchical detail simplification operations; the vertex level uses an improved edge folding algorithm, the patch level uses an adaptive deletion algorithm, and the region level uses a mesh reconstruction algorithm;

[0012] The simplified model is subjected to multi-dimensional quality assessment, and indices such as Hausdorff distance and volume deviation rate are calculated. If the indices do not meet the standards, the simplification rate is adjusted back and the model is simplified again.

[0013] Output a simplified model with multi-level details, store model data according to LOD level, optimize vertex sorting and texture storage, and record a simplified log.

[0014] Furthermore, it also includes introducing a feature importance weighted calculation in the multi-scale feature extraction step, with the feature importance value calculated using the formula... The calculation is performed, where I is the feature importance value and n is the number of features involved in the calculation. The weight coefficient for the k-th feature is... is the normalized value of the k-th feature.

[0015] Furthermore, it also includes incorporating a dynamic simplification rate adjustment mechanism into the simplified decision model. When the model's observation distance changes, the new objective simplification rate is adjusted using the formula... Calculate, where r is the new objective simplification rate. The initial simplification rate is the reference distance, and d is the current observation distance. This is the baseline distance.

[0016] Furthermore, in the edge folding algorithm, the folding cost is calculated comprehensively based on the curvature of the vertex, the area of ​​adjacent faces, and the amount of texture change. The curvature is represented by the average angle between the normal vectors of the faces surrounding the vertex. The area of ​​adjacent faces is calculated using the triangle area formula. The amount of texture change is determined by comparing the pixel differences in the texture mapping regions before and after folding. The coordinates of the new vertex generated after folding are a weighted average of the coordinates of the two endpoints of the folded edge. The weights are determined based on the curvature values ​​of the two endpoints, and the weight calculation formula is as follows: ,in , These are the weights of the two endpoints, , These are the curvature values ​​at the two endpoints.

[0017] Furthermore, in the quality assessment step, the volume deviation rate is determined by the formula... Calculation, where This refers to the volume deviation rate. To simplify the size of the model, Let be the volume of the original model, and let the integration range be the projection region of the model's bounding box in 3D space. The integration formula is: ,in The three-dimensional space occupied by the model.

[0018] Furthermore, in the region growth algorithm, the growth stopping condition is that the angle between the normal vector of the newly added vertex and the seed point is greater than 30 degrees or the distance is more than 3 times the average point spacing of the model. The average point spacing of the model is obtained by calculating the average distance between all vertices and their nearest neighbor vertices. After the region is divided, each region is assigned a unique region identifier, which includes the hash value of the region center coordinates and the region type code. The region identifier is associated with the feature parameters of the region and stored. The feature parameters include the average curvature of the region, the maximum height difference, the texture complexity, the number of triangles, and the semantic label. The texture complexity is calculated by the entropy value of the texture image.

[0019] Furthermore, it also includes adding a transition smoothing coefficient calculation during LOD level transitions. The transition smoothing coefficient is calculated using the formula... Calculation, where This is the transition smoothing coefficient. The transition start time, The transition end time, The transparency of the previous level model at time t. Let be the transparency of the next level model at time t, where the transparency changes linearly with time and satisfies... ,Right now , This coefficient is used to evaluate the transition effect.

[0020] Furthermore, in the texture compression process, the compressed texture resolution is determined by the formula... The calculation is performed, where R is the compressed texture resolution in pixels. The original texture resolution is in pixels, and S is the simplified area of ​​the current patch. This represents the original area of ​​the patch.

[0021] Furthermore, in the mesh reconstruction algorithm, the number of vertices in the reconstructed mesh is determined by the formula... Calculate, where N is the number of reconstructed vertices. Let r be the number of vertices before reconstruction, r be the simplification rate of the region, and the complexity coefficient be calculated based on the average curvature of the vertices within the region. The higher the curvature, the larger the coefficient. The specific calculation formula is as follows: Where c is the average curvature of the region. This represents the maximum curvature of the entire model.

[0022] Furthermore, in the multi-factor weighted algorithm, the priority is simplified using the formula... Calculate, where P is the simplification priority. It is a distance factor, positively correlated with the observation distance, through... Calculate, where d is the current observation distance. For the maximum distance of the model, The accuracy factor is negatively correlated with the accuracy requirements of the application scenario. Calculate, where t is the precision threshold for this region. The threshold for the maximum accuracy of the model. It is a hardware factor and is positively correlated with hardware processing power.

[0023] Compared with existing technologies, the beneficial effects of this invention are:

[0024] This invention lays a solid foundation for model simplification through refined preprocessing, effectively solving the imbalance problem of traditional preprocessing. Adaptive Gaussian filtering dynamically adjusts the kernel size based on the vertex neighborhood patch size, preserving details in dense regions while ensuring thorough noise removal in sparse regions. The improved RANSAC algorithm accurately removes outliers through multiple iterations, avoiding the accidental deletion of effective shape features. Block texture compression combines the patch projection area to determine the compression ratio; small patches use a high compression ratio to save memory, while large patches maintain high resolution to avoid distortion. Simultaneously, texture mapping coordinates are corrected to eliminate stretching issues, providing high-quality data support for subsequent simplification operations.

[0025] This invention significantly improves the rationality and relevance of simplification through multi-dimensional feature extraction and intelligent decision-making. Multi-scale feature extraction covers three levels: vertices, patches, and regions. Combined with visibility factors and semantic tags, it accurately identifies key areas such as building edges and vegetation outlines. The simplification decision-making comprehensively considers observation distance, scene accuracy, hardware performance, and semantic features. It sets a minimum simplification threshold to preserve structure for buildings and dynamically adjusts the simplification rate for vegetation and terrain. This meets the needs of high-precision scenarios such as building surveying while also adapting to efficient rendering for macro-planning. Furthermore, it matches the simplification intensity based on hardware scores, avoiding performance waste on high-end devices and lag on low-end devices, achieving adaptive adaptation to different scenarios and hardware.

[0026] This invention ensures the quality of model simplification and rendering effects through hierarchical simplification operations and multi-dimensional quality assessment. An improved edge-folding algorithm calculates folding costs based on semantic importance, prioritizing the preservation of key vertices such as building corners. Adaptive patch deletion and region mesh reconstruction ensure topological continuity of the simplified model; mesh merging in flat areas improves efficiency, while mesh optimization in complex areas preserves details. Multi-dimensional quality assessment covers distance, volume, texture, mesh quality, and visual similarity, comprehensively controlling the simplification effect and allowing for backtracking adjustments when requirements are not met to avoid local distortion. LOD level transitions optimize transparency changes through smoothing coefficients, eliminating abrupt transitions. Simplified logging records key parameters, supporting process traceability and iterative optimization. Overall, it achieves a balance between simplification efficiency, visual effects, and application requirements, significantly enhancing the application value of oblique photogrammetry-based 3D models across various fields. Attached Figure Description

[0027] Figure 1This is a schematic block diagram of an automatic simplification method for multi-level details of a real-world 3D model proposed in this invention;

[0028] Figure 2 A line graph showing the relationship between LOD levels and the number of triangles;

[0029] Figure 3 A line graph showing the relationship between observation distance and simplification rate;

[0030] Figure 4 A bar chart comparing vertex noise values ​​before and after preprocessing;

[0031] Figure 5 A bar chart comparing the quality assessment indicators of traditional methods and the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0034] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0035] Reference Figures 1 to 5 A method for automatically simplifying multi-level details of oblique photogrammetry real-world 3D models, comprising:

[0036] Obtain raw oblique photogrammetry data and basic information of the 3D model. The raw data includes multi-view oblique images, laser point clouds, POS data, and camera intrinsic parameters. The multi-view oblique images must meet the following requirements: forward overlap ≥80% and lateral overlap ≥70%. The density of the laser point cloud is set to 5-20 points / square meter according to the modeling accuracy requirements. The POS data includes the shooting coordinates and attitude angles of each image. The basic information of the 3D model includes vertex 3D coordinates, triangle patch indexes, texture coordinate mapping relationships, initial LOD level division, and model bounding box parameters. The raw data and the basic information of the 3D model are stored in association with spatial coordinates. A globally unique identifier and data source mapping table are established. The mapping table includes data acquisition time, equipment model, and accuracy level.

[0037] The 3D model is preprocessed, and an adaptive Gaussian filtering algorithm is used to denoise the vertex coordinates. The size of the filtering kernel is dynamically adjusted according to the average size of the vertex's neighboring face. The larger the size of the neighboring face, the larger the filtering kernel radius. The RANSAC algorithm is improved to remove outliers from the laser point cloud. The threshold for outlier determination is 1.5 times the average point spacing of the model. Point cloud data that are determined to be outliers in three consecutive iterations will be permanently removed. The texture image is processed by block compression. The compression ratio is determined based on the ratio of the projected area of ​​the texture block to the overall surface area of ​​the model. Texture blocks with a projected area ratio of less than 0.01% use a higher compression ratio. Simultaneously, the texture mapping coordinates are corrected by bilinear interpolation to eliminate texture stretching distortion caused by model simplification. The model topology is repaired by removing duplicate vertices, non-manifold edges, and degenerate faces. The criterion for degenerate faces is that the shortest side length is less than 1 / 10 of the longest side length.

[0038] Multi-scale features of the 3D model are extracted. At the vertex level, the normal vector, Gaussian curvature, average curvature, and average distance to k-neighboring vertices of each vertex are calculated. The value of k is 1.2 times the average number of vertex connections in the model. The Gaussian curvature is calculated by the rate of change of the normal vector of the vertex's adjacent face. At the face level, the face area, shape factor, slope, and visibility factor are calculated. The shape factor is the ratio of the face area to the area of ​​the smallest circumcircle. The visibility factor is determined by the ratio of the number of times the face is visible in all photographic views to the total number of views. At the region level, a region growing algorithm based on feature clustering is used to divide the model regions. Seed points are preferentially selected vertices with curvature greater than a set threshold and visibility factors higher than 0.7. During the growth process, the angle between the normal vector of the newly added vertex and the seed point must be less than 25 degrees, and the distance must be less than 2.5 times the average point spacing of the model. The volume, surface area, vertex density, average texture complexity, and semantic label of each region are calculated. The semantic label is obtained by deep learning model recognition and includes categories such as buildings, vegetation, roads, and water bodies.

[0039] A simplified decision-making model is constructed. The target simplification rate is determined by comprehensively considering the model observation distance, application scenario accuracy requirements, hardware performance parameters, and regional semantic features. The observation distance and the target simplification rate are mapped through a piecewise function. The simplification rate growth slope is 0.01 / meter within the range of 0-50 meters, 0.005 / meter within the range of 50-200 meters, and 0.002 / meter above 200 meters. The application scenario accuracy requirements are set with differentiated thresholds according to the region type. The position deviation threshold for building facades is set to 0.1 meters, for terrain surfaces to 0.5 meters, and for vegetated areas to 1.0 meter. Hardware performance parameters include graphics card memory capacity, CPU frequency, and memory bandwidth. The correspondence between hardware scores and the maximum number of triangles that can be supported is established through performance benchmark tests. The above factors are input into a multi-factor weighted algorithm to output the simplification priority of each region. Regions with higher priority are allowed to use higher simplification rates. At the same time, a minimum simplification rate threshold of 0.3 is set for regions with the semantic label "building" to retain key structures.

[0040] Hierarchical detail simplification operations are performed. For the vertex level, an improved edge folding algorithm is used. The folding cost is calculated by comprehensively considering vertex curvature, adjacent patch area, texture change, and semantic importance. Curvature accounts for 35%, patch area for 25%, texture change for 20%, and semantic importance for 20%. The semantic importance weight of building edge vertices is increased to 30%. The coordinates of the new vertex after folding are the weighted average of the coordinates of the two endpoints of the folded edge. The weight is positively correlated with the vertex curvature. For the patch level, an adaptive patch deletion algorithm is used. The deletion threshold is dynamically adjusted according to the ratio of patch area to the average patch area of ​​the region. Patches with a ratio less than 0.2 are marked as redundant patches. After deletion, the normal vector change rate of the remaining patches must be less than 15 degrees. For the region level, a simplification algorithm based on mesh reconstruction is used. Quadrilateral mesh merging is used for flat regions, and triangular mesh optimization is used for complex regions. During reconstruction, the continuity of the region boundary line is maintained, and the simplification rate of the boundary line vertices does not exceed 0.5. After the simplification operation, the vertex index, patch connection relationship, texture mapping data, and LOD level information are updated in real time.

[0041] The simplified model is evaluated for quality in multiple dimensions. The Hausdorff distance, volume deviation rate, texture matching degree, mean triangle aspect ratio, and visual similarity between the simplified model and the original model are calculated. The Hausdorff distance must be less than 1.2 times the accuracy threshold of the application scenario. The volume deviation rate is calculated by the model surface integral and controlled within 5%. The texture matching degree is determined by comparing the average pixel error of the texture projection before and after simplification. The error value must be less than 5 pixels. The mean triangle aspect ratio must be greater than 0.3 to ensure mesh quality. The visual similarity is evaluated by comparing the structural similarity index (SSIM) of the rendered images from the key viewpoint. The SSIM value must be greater than 0.9. If any evaluation index fails to meet the requirements, the simplification decision model is backtracked to adjust the simplification rate parameter of the corresponding region and the simplification operation is re-executed.

[0042] Output a multi-level detailed simplified model, storing model data with different simplification rates according to LOD levels. Each LOD level includes vertex cache data, patch index data, compressed texture atlas, level transition parameters, and simplification log. Vertex cache data is reordered using a vertex cache optimization algorithm to improve rendering efficiency. Texture atlases are merged according to region semantics, with textures from the same semantic regions stored together. Level transition parameters include transition distance thresholds, transparency change curves, and vertex position offsets, used to achieve smooth switching between different LOD levels during rendering. The simplification log records the simplification rate, operation time, and quality evaluation results for each region, supporting the traceability and optimization of the model simplification process.

[0043] This invention also includes introducing a feature importance weighted calculation in the multi-scale feature extraction step, whereby the feature importance value is calculated using a formula. The calculation is performed, where I is the feature importance value, and n is the number of features involved in the calculation, ranging from 5 to 8. Here, represents the weight coefficient for the k-th feature. The weight coefficient for the building edge feature is set to 0.2-0.3, and the weight coefficient for the flat area feature is set to 0.05-0.1. This is the normalized value of the k-th feature, with a normalization range of 0 to 1. This weighted calculation can prioritize the preservation of important feature regions when simplifying decisions, avoiding the loss of key details during the simplification process.

[0044] This invention also includes a dynamic simplification rate adjustment mechanism in the simplified decision model. When the model observation distance changes, the new target simplification rate is adjusted using the formula... Calculate, where r is the new objective simplification rate, ranging from 0.1 to 0.9. The initial simplification factor is 0.2 at the reference distance, and d is the current observation distance in meters. The baseline distance is 1 / 5 of the diagonal length of the model bounding box, in meters. This dynamic adjustment mechanism makes the simplification rate increase exponentially with the observation distance, significantly improving rendering efficiency while ensuring the visual effect of observations at a distance.

[0045] In this invention, the edge folding algorithm calculates the folding cost by comprehensively considering the curvature of the vertex, the area of ​​adjacent faces, and the amount of texture change. Curvature is represented by the average angle between the normal vectors of the faces surrounding the vertex; a larger angle indicates higher curvature. The area of ​​adjacent faces is calculated using the triangle area formula. The amount of texture change is determined by comparing the pixel differences in the texture mapping regions before and after folding. In the folding cost formula, curvature accounts for 40%, face area for 30%, and texture change for 30%. Folding is performed when the folding cost is less than a set threshold. The folding threshold is dynamically set based on the current LOD level of the model; higher levels result in larger thresholds. The coordinates of the newly generated vertex after folding are a weighted average of the coordinates of the two endpoints of the folded edge. The weights are determined based on the curvature values ​​of the endpoints; vertices with higher curvature have greater weights to preserve detailed features. The weight calculation formula is as follows: ,in , These are the weights of the two endpoints, , These are the curvature values ​​at the two endpoints.

[0046] In this invention, the volume deviation rate in the quality assessment step is determined by the formula... Calculation, where This refers to the volume deviation rate. To simplify the size of the model, The volume of the original model is calculated by performing a triple integral over the model's surface. The integration range is the projected area of ​​the model's bounding box in 3D space. The integration formula is: ,in For the three-dimensional space region occupied by the model, when the volume deviation rate exceeds 5%, local refinement processing is performed on the region with large deviation. The deviation is reduced by increasing the number of vertices in the region. The increase in the number of vertices is positively correlated with the volume deviation rate. For every 1% increase in deviation, the number of vertices in the region increases by 5% of the total number of vertices in the region.

[0047] In this invention, the region growth algorithm stops growth when the angle between the normal vector of the newly added vertex and the seed point is greater than 30 degrees or the distance is more than three times the average point spacing of the model. The average point spacing of the model is obtained by calculating the average distance between all vertices and their nearest neighbors. After region division, each region is assigned a unique region identifier. The region identifier is encoded using a 64-bit integer and includes the hash value of the region center coordinates and the region type code. The region identifier is associated with the feature parameters of the region. The feature parameters include the region's average curvature, maximum height difference, texture complexity, number of triangles, and semantic label. Texture complexity is calculated using the entropy value of the texture image. The higher the entropy value, the more complex the texture. In subsequent simplification operations, the corresponding simplification strategy is called according to the region identifier. The building region adopts the edge-preserving simplification strategy, and the vegetation region adopts the contour-preserving simplification strategy.

[0048] This invention also includes incorporating a transition smoothing coefficient calculation during LOD level transitions. The transition smoothing coefficient is calculated using the formula... Calculation, where This is the transition smoothing coefficient, with a value ranging from 0 to 0.25. The transition start time, The transition end time is set to a total transition duration of 0.5 seconds. The transparency of the previous level model at time t. Let be the transparency of the next level model at time t, where the transparency changes linearly with time and satisfies... ,Right now , This coefficient is used to evaluate the transition effect. When the coefficient is less than 0.8, the rate of change of transparency is adjusted and the transition time is extended to 0.8 seconds to improve smoothness.

[0049] In this invention, the texture compression process determines the compressed texture resolution using the formula... The calculation is performed, where R is the compressed texture resolution in pixels. The original texture resolution is in pixels, and S is the simplified area of ​​the current patch in square meters. The original patch area is expressed in square meters. This calculation matches the texture resolution with the patch area, avoiding memory waste caused by small patches corresponding to high-resolution textures. The compression process uses the ETC2 compression format, which supports the Alpha channel and has a fixed compression ratio of 1:4. Before compression, the texture image is subjected to Gamma correction. The correction coefficient is dynamically adjusted according to the lighting conditions during photography. The correction coefficient is set to 1.2 for areas with high lighting intensity and 0.8 for areas with low lighting intensity.

[0050] In this invention, the number of vertices in the reconstructed mesh is determined by the formula... Calculate, where N is the number of reconstructed vertices. Let r be the number of vertices before reconstruction, and r be the simplification rate of the region, ranging from 0 to 0.8. This is the region complexity coefficient, ranging from 0.5 to 1.5. The complexity coefficient is calculated based on the average curvature of the vertices within the region; the higher the curvature, the larger the coefficient. The specific calculation formula is as follows: Where c is the average curvature of the region. This is the maximum curvature of the model as a whole. This calculation ensures that more vertices are retained in complex regions, thus maintaining the structural characteristics of the model.

[0051] In this invention, the priority is simplified using the formula in the multi-factor weighted algorithm. Calculate, where P is the simplification priority, ranging from 0 to 1. This is a distance factor, positively correlated with the observation distance; the greater the observation distance, the larger the value. Calculate, where d is the current observation distance. The maximum visible distance of the model. This is a precision factor, negatively correlated with the precision requirements of the application scenario; the lower the precision requirement, the larger the value. Calculate, where t is the precision threshold for this region. The threshold for the maximum accuracy of the model. The hardware factor is positively correlated with hardware processing power; the stronger the processing power, the larger the value. It is obtained through hardware scoring normalization. All factors are normalized to the range of 0 to 1. Regions with higher priority are simplified with a higher simplification rate, with the highest simplification rate reaching 0.8, in order to maximize the overall processing efficiency while ensuring visual effects.

[0052] The following two examples further illustrate the specific implementation of this system:

[0053] Example 1: Simplified Building-Dense Scenes in the Core Area of ​​a Smart City Using Oblique Photogrammetry Model

[0054] This embodiment takes the core area of ​​a smart city in a first-tier city as the application object. The area covers 5km², including 80 high-rise buildings, 200 multi-story buildings, 30km of roads, and a small amount of vegetation and water. The oblique photography data was collected using a DJI P1 full-frame oblique camera, and the laser point cloud was acquired by a RIEGLVZ-6000 3D laser scanner. It is necessary to generate a LOD0-4 five-level detail model to adapt to scenarios such as urban planning approval and emergency command. The focus is on verifying semantic-driven simplification, multi-factor decision-making, and texture adaptation capabilities. The feature importance formula, dynamic simplification rate formula, and LOD transition smoothness coefficient formula are applied simultaneously.

[0055] 1. Detailed Implementation of Core Steps

[0056] Step 1: Acquisition of raw data and basic information. The oblique image has a lateral overlap of 85% and a single image resolution of 8192×5460 pixels, a focal length of 35mm, and a total of 2000 images are acquired. The laser point cloud density is 15 points / square meter, and the point cloud coordinate accuracy is ±2cm, including X / Y / Z three-dimensional coordinates and reflection intensity information. The POS data is acquired by the D-RTK2 high-precision positioning system. The shooting coordinate error of each image is ≤5cm, and the attitude angle roll, pitch, and heading errors are ≤0.1°. The camera intrinsic parameters are obtained through camera calibration. The focal length is f=34.98mm, the principal point coordinates are 4096 pixels and 2730 pixels, and the distortion coefficients are k1=-0.012 and k2=0.003. The basic information of the 3D model was generated by ContextCapture modeling software. The initial model contains 120 million vertices and 240 million triangles, with a texture resolution of 4096×4096 pixels per block. The initial LOD levels are divided according to the number of triangles: LOD0 has 240 million faces, LOD1 has 120 million faces, LOD2 has 60 million faces, LOD3 has 30 million faces, and LOD4 has 15 million faces. The model bounding box parameters are: X-axis 324500m to 325500m, Y-axis 512000m to 513000m, Z-axis 0m to 300m, and diagonal length 1414m. All data is stored in association with spatial coordinates. The global identifier uses the encoding of "image ID-point cloud block ID-model face ID". The mapping table records the data acquisition time as 2024-05-10, the equipment model, and the accuracy level as millimeters.

[0057] Step 2: Model Preprocessing Adaptive Gaussian Filtering: The average size of the vertex neighborhood patches in the building region is 0.2m, and the filter kernel radius is set to 0.5; the average size of the terrain patches is 1.0m, and the filter kernel radius is set to 1.2. After filtering, the coordinate deviation of building corner vertices is ≤0.5cm, and the deviation of terrain vertices is ≤2cm. An improved RANSAC algorithm is used to remove outliers from the point cloud. The algorithm iterates 3 times: the threshold for the first iteration is 1.5 times the average point spacing (0.3m), the threshold for the second iteration is 0.3m, and the threshold for the third iteration is 0.2m. The final outlier removal rate is 1.2%, while retaining effective features such as terrain protrusions like artificial mountains. Texture Block Compression: The projected area of ​​building facade patches accounts for 0.05%, with a compression ratio of 1:8; the projected area of ​​road patches accounts for 0.2%, with a compression ratio of 1:4; the projected area of ​​vegetation patches accounts for 0.01%, with a compression ratio of 1:16. The compression format uses ETC2, supporting an Alpha channel to achieve water transparency effects. Bilinear interpolation corrects texture mapping coordinates, ensuring building facade texture stretching error ≤ 1 pixel and road texture stretching error ≤ 3 pixels. Topology repair: Removes 320,000 duplicate vertices, 150,000 non-manifold edges, and 80,000 degenerate faces. The criteria for degenerate faces are shortest side < 0.05m and longest side > 0.5m.

[0058] Step 3: Multi-scale Feature Extraction Vertex Level: k value is set to 12, average vertex connectivity is 10, Gaussian curvature is calculated by the angle between the normal vectors of 3 adjacent patches, Gaussian curvature of building corner points > 0.8, wall vertices < 0.2; average k-neighborhood distance is 0.1m for building areas and 0.5m for terrain areas. Patch Level: Building facade patch area 0.01~0.1m², shape factor 0.8-0.9, close to an equilateral triangle; terrain patch area 0.5~2m², shape factor 0.3-0.6; visibility factor is calculated by counting the number of times the patch is visible in 2000 images, building facade visible more than 1000 times, visibility factor > 0.5, terrain patch visible more than 500 times, visibility factor > 0.25. Regional Hierarchy: Semantic labels are identified based on the ResNet50 semantic segmentation model, with an accuracy of 98% for buildings, 96% for roads, 95% for vegetation, and 97% for water bodies. Seed points for regional growth are selected from building corners with curvature > 0.6 and visibility factor > 0.7. During growth, the angle between normal vectors is ≤ 25° and the distance is ≤ 0.5m. The average point spacing is 0.2m multiplied by 2.5. A total of 5000 regions are divided. The volume of each region is calculated, with building regions ranging from 50 to 5000m³, as well as surface area, vertex density (50 points / m² for building regions and 10 points / m² for terrain regions), and texture entropy value (> 8 for building facades and < 4 for water bodies).

[0059] Step 4: Simplify the decision-making model and construct observation distance mappings: 0-50m is the walking view, with a simplification rate of 0.1-0.3; 50-200m is the vehicle view, with a simplification rate of 0.3-0.6; and above 200m is the drone view, with a simplification rate of 0.6-0.9. Scene accuracy thresholds: building facade position deviation ≤0.1m, meeting planning approval requirements; terrain ≤0.5m, meeting macro-planning requirements; vegetation ≤1.0m, meeting landscape display requirements. Hardware performance benchmark tests: RTX4090 graphics card with 24GB VRAM, maximum triangle capacity of 80 million faces; i7-13700K CPU with a base frequency of 5.4GHz, processing speed of 20 million faces / second, hardware score of 100 points; mid-range laptop with MX550 graphics card, maximum capacity of 20 million faces, hardware score of 30 points. Multi-factor weighting: building area simplification priority. , For distance factor, For the precision factor, For hardware factors, such as when the observation distance is 200m =0.5, building accuracy threshold 0.1m =0.2, high-end hardware =1.0, then P=0.4×0.5+0.3×0.2+0.3×1.0=0.2+0.06+0.3=0.56, corresponding to a simplification rate of 0.6; the minimum simplification rate for the vegetation area is 0.8, P=0.4×0.8+0.3×0.9+0.3×1.0=0.32+0.27+0.3=0.89, corresponding to a simplification rate of 0.9.

[0060] Step 5: Hierarchical Simplification of Vertex Level: Improve the edge folding algorithm. Folding cost = 0.35 × curvature + 0.25 × patch area + 0.2 × texture change + 0.2 × semantic importance. Semantic importance for building corners is 0.3, and for wall vertices it is 0.1. The folding threshold from LOD0 to LOD1 is 0.2, and from LOD1 to LOD2 it is 0.3. The new vertex coordinate weight w1 = c1 / (c1 + c2), where w1 and w2 are the weights of the two endpoints, and c1 and c2 are the curvature values ​​of the two endpoints. For example, for building corners, c1 = 0.9; for wall vertices, c2 = 0.1; w1 = 0.9, w2 = 0.1; retaining the corner position. Patch Level: Adaptive deletion threshold = average patch area of ​​the region × 0.2. The average patch area for a building region is 0.05m², with a deletion threshold of 0.01m². After deletion, the normal vector change rate is ≤15°. The average patch area for a terrain region is 1m², with a deletion threshold of 0.2m². Region hierarchy: Building areas are optimized using triangular meshes, retaining edge lines with a simplification rate of ≤0.5; terrain areas are merged using quadrilateral meshes, with 10 small faces merged into 1 large face; after simplification, vertex indices are updated, GPU vertex cache sorting is used, and sorting is done in ascending order of Z coordinates to improve rendering efficiency, as well as face connection relationships, redundant indices are removed, and texture mapping is updated to the merged face.

[0061] Step Six: Quality Assessment and Formula Application - Feature Importance Calculation: Building area n=6, where n is the number of features involved in the calculation, including curvature, area, visibility, semantics, texture entropy, and distance. ω1 is the curvature weight coefficient = 0.25, ω2 is the area weight coefficient = 0.15, ω3 is the visibility weight coefficient = 0.2, ω4 is the semantic weight coefficient = 0.2, ω5 is the texture entropy weight coefficient = 0.1, and ω6 is the distance weight coefficient = 0.1; f1 is the curvature normalized value = 0.9 (high curvature), and f2 is the area normalized value = 0.5 (medium area). f3 is the visibility normalization value = 0.8, indicating high visibility; f4 is the semantic normalization value = 1.0, indicating architectural semantics; f5 is the texture entropy normalization value = 0.9, indicating high entropy texture; f6 is the distance normalization value = 0.5, indicating medium distance. Substituting these values ​​into the formula I = 0.25 × 0.9 + 0.15 × 0.5 + 0.2 × 0.8 + 0.2 × 1.0 + 0.1 × 0.9 + 0.1 × 0.5 = 0.225 + 0.075 + 0.16 + 0.2 + 0.09 + 0.05 = 0.8, where I is the feature importance value, and higher I values ​​are prioritized. Dynamic simplification rate: d = 200m, which is the current observation distance; d0 = 1414 / 5 = 282.8m, which is 1 / 5 of the bounding box diagonal; r0 = 0.2, which is the baseline simplification rate. Substituting these values ​​into the formula... 'r' represents the new target simplification rate, corresponding to LOD2 level, i.e., 60 million faces. LOD transition smoothness coefficient: t0 = 0s is the transition start time, t1 = 0.5s is the transition end time, and the total transition duration is set to 0.5 seconds. α(t) is the transparency of the previous level model at time t, and β(t) is the transparency of the next level model at time t. The transparency changes linearly with time and satisfies α(t) + β(t) = 1, i.e., α(t) = 1 - 2t, β(t) = 2t. Substituting these values ​​into the formula... τ is the transition smoothness coefficient. τ = 0.083 < 0.8. Extend the transition time to 0.8 seconds, α(t) = 1 - 1.25t, β(t) = 1.25t. Recalculate τ = 0.128 > 0.8, and there is no transition jump.

[0062] 2. Effect Verification and Table Analysis

[0063] Table 1 Comparison of Simplified Quality of Smart City Core Area Models

[0064]

[0065] Table 1 shows the data based on comparative tests of five typical areas in the core area of ​​a smart city, fully demonstrating the quality advantages of this invention. Traditional simplification methods do not consider semantic features; the Hausdorff distance of the building area (0.15m) exceeds the 0.1m threshold for planning approval, and the volume deviation rate of 8.2% leads to errors in building volume ratio calculation. This invention, through semantic-driven simplification, improves building area accuracy by 46.7%, and controls volume deviation within 3.5%, meeting the dimensional accuracy requirements of planning approval. Texture matching accuracy is improved to 0.93, avoiding the blurring of building facades caused by texture stretching in traditional methods; triangle aspectratio is improved to 0.38, reducing face distortion during rendering and lowering the GPU rendering burden. The LOD transition jump rate is reduced from 28% to 3%, and transparency changes are optimized through a smoothing coefficient, resulting in no noticeable stuttering during viewpoint switching in emergency command scenarios, improving the real-time interactive experience and verifying the effectiveness of semantic simplification, multi-factor decision-making, and transition optimization.

[0066] Example 2: Mountainous terrain oblique photogrammetry model simplifies terrain-dominant scenes

[0067] This embodiment uses a mountainous area of ​​a provincial nature reserve as the application object, with an area of ​​50 km², including 30 km² of mountain slopes greater than 25°, 15 km² of valley slopes of 5-25°, and 5 km² of water bodies, i.e., lakes. There are no dense buildings. The oblique imagery was collected by a Wing Loong-2 mapping fixed-wing UAV, and the laser point cloud was acquired by a LiDAR sensor. A four-level model of LOD0-3 needs to be generated to adapt to ecological monitoring and forest fire prevention scenarios. The focus is on verifying the efficiency of terrain simplification, LOD transition, and hardware adaptability. The formulas for volume deviation rate, texture resolution, and number of mesh vertices are applied simultaneously.

[0068] 1. Detailed Implementation of Core Steps

[0069] Step 1: Acquisition of raw data and basic information: The oblique image has a forward overlap of 80% and a lateral overlap of 70%. The resolution of a single image is 5472×3648 pixels, the focal length is 50mm, and a total of 1500 images are acquired. The laser point cloud density is 5 points / square meter, the point cloud coordinate accuracy is ±5cm, and the reflection intensity is used to distinguish vegetation from rocks. The POS data is acquired by the Beidou high-precision positioning system, with a shooting coordinate error ≤10cm and an attitude angle error ≤0.2°. The camera intrinsic parameters are: focal length f=49.97mm, principal point coordinates are 2736 pixels and 1824 pixels, and distortion coefficients k1=-0.01 and k2=0.002. The basic information of the 3D model was generated by Pix4Dmapper. The initial model contains 80 million vertices and 160 million triangles, with a texture resolution of 2048×2048 pixels per block. The initial LOD levels are: LOD0 with 160 million faces, LOD1 with 80 million faces, LOD2 with 40 million faces, and LOD3 with 20 million faces. The model bounding box parameters are: X-axis 123000m to 128000m, Y-axis 456000m to 461000m, Z-axis 500m to 1500m, and diagonal length 7071m. The data is stored in association, and the global identifier includes "point cloud block ID - terrain zone ID". The mapping table records the acquisition time as 2024-06-01, the device model, and the accuracy level as centimeter-level.

[0070] Step 2: Model Preprocessing Adaptive Gaussian Filtering: The average size of the mountain area patches is 1.5m, and the filter kernel radius is 1.8; the average size of the valley area patches is 0.8m, and the kernel radius is 1.0; after filtering, the deviation of mountain vertices is ≤5cm, and the deviation of valley vertices is ≤3cm. An improved RANSAC algorithm is used to remove outliers from the point cloud, iterated three times, with thresholds of 0.75m (1.5 × average point spacing of 0.5m, 0.6m, and 0.45m), achieving an outlier removal rate of 2.5%, while retaining effective features such as rock protrusions and tree roots. Texture Block Compression: The projected area of ​​the mountain area patches accounts for 0.5%, with a compression ratio of 1:4; the valley area patches account for 0.3%, with a compression ratio of 1:6; the water area patches account for 0.1%, with a compression ratio of 1:12; the compression format is ETC2, and the Gamma correction coefficients are set as follows: mountain illumination intensity coefficient is set to 1.2, and valley illumination weakness coefficient is set to 0.8. Topology repair: Removed 180,000 duplicate vertices, 90,000 non-manifold edges, and 50,000 degenerate faces. The shortest side of the degenerate face is less than 0.1m and the longest side is greater than 1m.

[0071] Step 3: Multi-scale feature extraction. Vertex level: k value is 8, average vertex connectivity is 7, Gaussian curvature is >0.5 for mountains (rocky areas, <0.3 for valleys); average k-neighborhood distance is 0.6m for mountains and 0.3m for valleys. Patch level: Mountain patch area is 1-3m², shape factor 0.3-0.5; valley patch area is 0.5-1m², shape factor 0.5-0.7; visibility factor: mountain patches are visible more than 300 times with a total viewing angle of 1500, and valley patches are visible more than 500 times. Regional hierarchy: Semantic labels are identified by a semantic segmentation model, with an accuracy of 90% for rocks, 92% for vegetation, and 98% for water. Seed points for regional growth are selected from rock vertices with curvature > 0.4 and visibility factor > 0.6. During growth, the angle between normal vectors is ≤ 25°, the distance is ≤ 1.25m, and the average point spacing is 0.5m multiplied by 2.5. A total of 3000 regions are divided. The volume is calculated, with the volume of mountain regions ranging from 1000 to 10000m³, as well as the surface area, vertex density (8 points / m² for mountains and 12 points / m² for valleys), and texture entropy values ​​(rock entropy > 6 and water entropy < 3).

[0072] Step 4: Simplify the decision-making model and construct observation distance mappings: 0-50m for foot monitoring (simplification rate 0.1-0.2), 50-200m for vehicle patrol (simplification rate 0.2-0.5), 200-500m for drone patrol (simplification rate 0.5-0.8), and above 500m for satellite view (simplification rate 0.8-0.95); Scene accuracy thresholds: Mountain terrain deviation ≤ 0.5m, meeting ecological monitoring needs; water boundary deviation ≤ 1.0m, meeting firebreak planning needs. Hardware performance benchmark testing: The forest fire command vehicle is equipped with an RTX 3060 graphics card with 12GB of VRAM, a maximum capacity of 40 million polygons, and a hardware score of 60; the computers at grassroots stations are equipped with GTX 1650 graphics cards with 4GB of VRAM, a maximum capacity of 10 million polygons, and a hardware score of 20. Multi-factor weighting: Mountainous areas For distance factor, For the precision factor, For hardware factors, such as when the observation distance is 500m =0.9, terrain accuracy threshold 0.5m =0.5, Command vehicle hardware time =0.6, then P=0.4×0.9+0.3×0.5+0.3×0.6=0.36+0.15+0.18=0.69, corresponding to a simplification rate of 0.8; the minimum simplification rate for the water body area is 0.9, P=0.4×0.95+0.3×0.9+0.3×0.6=0.38+0.27+0.18=0.83, corresponding to a simplification rate of 0.95.

[0073] Step 5: Hierarchical Simplification Operation Vertex Level: Edge folding cost = 0.35 × curvature + 0.25 × patch area + 0.2 × texture change + 0.2 × semantic importance, with a semantic importance of 0.25 for rock vertices and 0.08 for vegetation vertices; LOD0→LOD1 folding threshold 0.25, LOD1→LOD2 threshold 0.35; New vertex weights are allocated according to curvature, with c1=0.6 for rock vertices, c2=0.2 for vegetation vertices, w1=0.75, w2=0.25, where w1 and w2 are the weights of the endpoints, and c1 and c2 are the curvature values ​​of the endpoints. Patch Level: Deletion threshold = average patch area of ​​the region × 0.2, with an average patch size of 2m² for mountains and a threshold of 0.4m²; and an average patch size of 0.8m² for valleys and a threshold of 0.16m², with a normal vector change rate ≤ 15° after deletion. Regional hierarchy: Mountains are merged using quadrilateral meshes, with 5 faces merged into 1; valleys are optimized using triangular meshes to preserve terrain undulations; after simplification, vertex caches are sorted by X / Y coordinates, terrain rendering is prioritized by planar position, texture atlases are merged by terrain type, and rock textures are stored centrally.

[0074] Step Six: Quality Assessment and Formula Application Calculation of Volume Deviation Rate: Original Model Volume =50000m³, representing the volume of a mountainous region, simplified model. =48250m³, substitute into the formula δ is the volume deviation rate; 3.5% < 5%, which meets the requirement. If δ = 6.2% for a certain valley region, then increase the number of vertices in that region. The total number of vertices is 100,000 multiplied by 5% and then by 1.2, which equals 6,000. After simplification, δ = 4.8%. Texture resolution calculation: Original texture R0 = 2048 pixels, R0 is the original texture resolution; original patch S0 = 2m², S0 is the area of ​​the original patch; after simplification, patch S = 1.6m², S is the area of ​​the current simplified patch. Substitute into the formula... Pixels, R is the compressed texture resolution, which is 1831 pixels, matching the area of ​​the facet, without stretching. Mesh vertex count calculation: N0 = 100,000 before reconstruction, representing the number of vertices in a mountainous region before reconstruction; simplification rate r = 0.8; average curvature of the region c = 0.5; maximum curvature of the model. =1.0, γ=0.5+1.0×(0.5 / 1.0)=1.0, where γ is the domain complexity coefficient. Substituting into the formula... 10,000, where N is the number of reconstructed vertices. The reconstructed vertices number 20,000, preserving the rock undulation features.

[0075] 2. Effect Verification and Table Analysis

[0076] Table 2 Comparison of Simplification Efficiency and Visual Effect of Mountainous Terrain Model

[0077]

[0078] Table 2 shows data based on tests in five terrain zones of mountainous areas, highlighting the advantages of this invention in terms of efficiency and adaptability. Traditional simplification methods require 120 minutes to process a 160 million-face model and consume 8.2GB of video memory, which is insufficient for basic station computers with only 4GB of video memory. This invention, through terrain simplification and texture adaptation, reduces the processing time to 45 minutes, lowers LOD2 level video memory to 5.1GB, and allows for smooth loading on basic equipment. The rendering frame rate is increased to 45fps, meeting the interactive needs of real-time forest fire prevention patrols; the terrain undulation restoration accuracy is improved to 96%, with no loss of details in rocky areas, enabling accurate identification of vegetation cover changes in ecological monitoring. Hardware compatibility has been expanded from traditional mid-to-high-end devices to low-end devices, solving the problem of insufficient computing power at basic stations. This verifies the efficiency of terrain simplification, texture adaptation, and hardware adaptation, meeting the practical application needs of mountainous ecological monitoring and forest fire prevention.

[0079] Reference Figure 2 This chart visually illustrates the core logic of LOD (Level of Detail) hierarchy. LOD0, the highest level of detail, contains 240 million polygons, suitable for close-up observation (such as a walking perspective), clearly displaying the texture and detailed structure of building facades. As the level increases, the number of triangles decreases by approximately a 1:2 ratio, with LOD4 containing only 15 million polygons, suitable for long-distance observation (such as a drone's perspective). This step-by-step reduction aligns with the "hierarchical detail simplification" approach, ensuring high fidelity for close-up observations while reducing the hardware load for long-distance rendering by decreasing the number of polygons. The chart data validates the rationality of the LOD hierarchy, avoiding the problem of "blurry when viewed up close" or "stuttering when viewed from a distance" for a single model in different scenarios, and providing a foundation for subsequent dynamic simplification rate adjustments.

[0080] Reference Figure 3 This figure reflects the actual effect of the "dynamic simplification rate adjustment mechanism." The simplification rate increases with the observation distance and conforms to the formula. The model exhibits an exponential growth pattern. At a close distance of 50 meters, the simplification rate is only 0.3, retaining 70% of the model details and ensuring that key features such as building facades and terrain protrusions are not lost. At a distant distance of 500 meters, the simplification rate reaches 0.9, retaining only 10% of the details, significantly reducing the GPU rendering load. This dynamic adjustment balances visual effects and rendering efficiency—avoiding detail blurring caused by excessive simplification at close range while resolving stuttering issues caused by high-detail models at long distances, making it suitable for multi-view observation scenarios in smart cities, including walking, vehicle-mounted, and drone-based observations.

[0081] Reference Figure 4This figure verifies the effectiveness of the "model preprocessing step." Preprocessing, through adaptive Gaussian filtering and an improved RANSAC algorithm, significantly reduces vertex noise in each region. Building corner noise is reduced from 1.2cm to 0.3cm, ensuring accurate corner positions during subsequent edge folding and preventing structural deformation. Mountain rock noise is reduced from 2.5cm to 0.8cm, preserving topographic features such as rock protrusions, meeting the accuracy requirements of ecological monitoring. Noise values ​​in other regions are reduced by more than 60%, providing high-quality vertex data for multi-scale feature extraction (such as curvature and visibility factor calculation), avoiding feature misjudgments caused by noise, and laying the foundation for simplified decision-making in the future.

[0082] Reference Figure 5 This figure comprehensively demonstrates the advantages of this invention in quality assessment, covering the core indicators of "multi-dimensional quality assessment." The Hausdorff distance decreased from 0.15m to 0.08m, meeting the 0.1m accuracy threshold for building planning approval; the volume deviation rate decreased from 8.2% to 3.5%, avoiding volumetric calculation errors caused by volume distortion in traditional methods; the texture matching accuracy (SSIM) increased from 0.82 to 0.93, resolving visual blurring caused by texture stretching; and the triangle aspect ratio increased from 0.22 to 0.38, improving mesh quality and reducing face distortion during rendering. These data confirm the effectiveness of the "hierarchical simplified operation" and the "quality assessment backtracking mechanism," ensuring a balance between accuracy, visual effect, and rendering efficiency in the simplified model.

[0083] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic multi-level detail simplification of a photorealistic 3D model, characterized in that, The method comprises the following steps: Obtaining oblique photography raw data and three-dimensional model basic information; the raw data includes multi-view oblique images, laser point clouds, POS data and camera internal parameters, and the three-dimensional model basic information includes vertex coordinates, triangle face index, texture coordinate mapping relationship, initial LOD level division and model bounding box parameters; the two types of data are stored by associating spatial coordinates, a globally unique identifier mapping table is established, and data collection time, device model and precision level are recorded; Pretreating the three-dimensional model; adopting adaptive Gaussian filtering to denoise the vertex, improving the RANSAC algorithm to remove laser point cloud outliers, block compressing the texture image and correcting the mapping coordinates, repairing the model topology, and removing duplicate vertices, non-manifold edges and degenerate face plates; Extracting multi-scale features of the three-dimensional model; Calculating vertex level parameters, face level indicators, and dividing regions by using a feature clustering algorithm; Building a simplification decision model; Setting region differentiated thresholds and the minimum simplification rate by comprehensively observing the distance, application scene precision, hardware performance and semantic features, and outputting the simplification priority of each region by a multi-factor weighted algorithm; Performing hierarchical detail simplification operations; the vertex level uses an improved edge folding algorithm, the face level uses an adaptive deletion algorithm, and the region level uses a mesh reconstruction algorithm; Performing multi-dimensional quality evaluation on the simplified model, calculating Hausdorff distance and volume deviation rate indicators, and backtracking and adjusting the simplification rate and re-simplifying when the indicators do not meet the requirements; Outputting a multi-level detail simplification model, storing model data according to the LOD level, optimizing vertex ordering and texture storage, and recording simplification logs.

2. The method of claim 1, wherein the method further comprises: Also included in the multiscale feature extraction step is the introduction of feature importance weighting calculation, the feature importance value is calculated by the formula where I is the feature importance value, n is the number of features involved in the calculation, is the weight coefficient of the kth feature, is the normalized value of the kth feature.

3. The method of claim 1, wherein the method further comprises: Also included is a dynamic simplification rate adjustment mechanism in the simplified decision model. When the model observation distance changes, the new target simplification rate is calculated by the formula where r is the new target simplification rate, is the initial simplification rate under the reference distance, d is the current observation distance, is the reference distance.

4. The method of claim 1, wherein the method further comprises: In the edge collapse algorithm, the collapse cost is calculated by the curvature of the vertex, the area of the adjacent patch and the texture change amount, the curvature is represented by the average of the included angle of the normal vector of the patch around the vertex, the area of the adjacent patch is calculated by the triangular area formula, and the texture change amount is determined by comparing the pixel difference of the texture mapping area before and after the collapse; the new vertex coordinates generated after the collapse are the weighted average of the coordinates of the two endpoints of the collapsed edge, and the weight is determined according to the curvature value of the two endpoints, and the weight calculation formula is wherein , are the weights of the two endpoints respectively, , are the curvature values of the two endpoints respectively.

5. The method of claim 1, wherein the method further comprises: In the quality evaluation step, the volume deviation rate is calculated by the formula wherein is the volume deviation rate, is the volume of the simplified model, is the volume of the original model, and the integral range is the projection area of the model bounding box in the three-dimensional space, and the integral formula is wherein is the three-dimensional space area occupied by the model.

6. The method of claim 1, wherein the method further comprises: The feature clustering algorithm adopts a region growing algorithm, in which the growth stopping condition is that the angle between the normal vector of the new vertex and the seed point is greater than 30 degrees or the distance exceeds 3 times the average point distance of the model, the average point distance of the model is obtained by calculating the average distance between all vertices and their nearest neighbor vertices, a unique region identifier is assigned to each region after region division, including the hash value of the region center coordinates and the region type code, the region identifier is stored in association with the feature parameters of the region, and the feature parameters include the average curvature, maximum height difference, texture complexity, number of triangles and semantic label of the region, and the texture complexity is calculated by the entropy value of the texture image.

7. The method of claim 1, wherein the method further comprises: Also included is the addition of a transition smoothing factor calculation when transitioning between LOD levels, the transition smoothing factor being calculated by the formula where is the transition smoothing factor, is the transition start time, is the transition end time, is the previous level model transparency at time t, is the next level model transparency at time t, the transparency changing linearly over time and satisfying i.e. , This factor is used to evaluate the transition effect.

8. The method of claim 1, wherein the method further comprises: In the texture compression process, the compressed texture resolution is calculated by the formula , where R is the compressed texture resolution in pixels, is the original texture resolution in pixels, and S is the area of the current patch after simplification, is the area of the original patch.

9. The method of claim 1, wherein the method further comprises: In the mesh reconstruction algorithm, the number of reconstructed mesh vertices is calculated by the formula where N is the number of reconstructed vertices, is the number of vertices before reconstruction, r is the simplification rate of the region, and the complexity coefficient is calculated according to the average curvature of the vertices in the region, and the higher the curvature, the greater the coefficient. The specific calculation formula is where c is the average curvature of the region, is the maximum curvature of the entire model.

10. The method of claim 1, wherein the method further comprises: In the multi-factor weighting algorithm, the simplification priority is calculated by formula , wherein P is the simplification priority, is a distance factor positively correlated with the observation distance, calculated by , wherein d is the current observation distance, is a maximum distance of the model, is an accuracy factor negatively correlated with the accuracy requirement of the application scenario, calculated by , wherein t is the accuracy threshold of the region, is a maximum accuracy threshold of the model, is a hardware factor positively correlated with the hardware processing capability.

Images