A 3D model lightweight method and system based on point cloud layered slicing and dynamic map mapping

Abstract

This invention relates to the fields of computer graphics and 3D modeling optimization technology, specifically to a lightweight 3D model method and system based on point cloud layering and dynamic texture mapping. In use, this invention establishes a bidirectional coupling mechanism between point clouds and textures, thereby improving storage efficiency, rendering performance, and visual fidelity. Regarding storage efficiency, it relies on texture curvature adaptive segmentation and dynamic slicing binding technology to construct an efficient associative storage system for geometry and texture, achieving excellent geometry and texture compression effects and significantly reducing the storage cost of high-precision 3D models. In terms of real-time rendering performance, it significantly improves real-time model scheduling efficiency through runtime dynamic loading strategies combined with efficient visibility filtering acceleration and intelligent resource preloading technology, thereby significantly improving work efficiency in scenarios such as design iteration and model editing.

Description

Technical Field

[0001] This invention relates to the fields of computer graphics and 3D modeling optimization technology, specifically to a method and system for lightweighting 3D models based on point cloud layering and dynamic texture mapping. Background Technology

[0002] With the rapid development and widespread application of 3D digitization technology, the acquisition and processing of high-precision 3D models has become a core requirement in fields such as industrial inspection, digital twins, and cultural heritage protection.

[0003] However, traditional 3D model processing techniques have an inherent flaw of processing geometric data and texture data separately, which leads to serious challenges in terms of data storage efficiency, real-time rendering performance, visual fidelity, and dynamic updates.

[0004] The current technological bottlenecks are mainly reflected in the following aspects:

[0005] The improved accuracy of 1D scanning allows the data size of a single model to reach tens of gigabytes, resulting in more than 30% storage redundancy in traditional UV mapping storage methods.

[0006] Second, real-time rendering applications require a frame rate of 90fps or higher and a latency of <20ms, while traditional texture loading mechanisms cannot meet the hardware limitations of mobile devices.

[0007] Third, professional fields such as industrial inspection require the preservation of sub-millimeter level feature details, but existing compression technologies result in up to 60% loss of high-frequency information;

[0008] Fourth, dynamic modifications require a full update of the model data, and in scenarios such as automotive design, a single modification can take more than 45 minutes.

[0009] Existing solutions such as Unreal Nanite virtual texture technology support large-scale texture streaming, but rely on dedicated graphics cards and cause a surge in power consumption of 300-400mW on mobile devices. Point cloud direct rendering solutions (PCL library) avoid geometric simplification losses, but lack material expression capabilities and have low rendering efficiency (<10fps @1M points). Neural rendering methods (NeRF) can achieve high-fidelity view composition, but have inherent limitations such as high training costs (8 V100s are required for 12 hours of training per scene) and lack of support for real-time editing. None of these technical approaches have fundamentally solved the problem of synergistic optimization between geometry and texture processing.

[0010] In summary, traditional 3D model processing techniques typically process and store geometric information (point cloud / mesh) and texture information separately. This separate approach has many limitations in terms of efficiency, accuracy, and resource utilization. Summary of the Invention

[0011] This invention establishes a two-way coupling mechanism between point clouds and textures, thereby improving storage efficiency, rendering performance, and visual fidelity, and providing a solution for lightweight processing of 3D models.

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] This invention provides a lightweight 3D model method based on point cloud layering and dynamic texture mapping, comprising the following steps:

[0014] S1. Data acquisition and preprocessing;

[0015] A hybrid acquisition system consisting of a phase-detection laser scanner and a high dynamic range texture camera was used for synchronous acquisition of multimodal data.

[0016] S2, Point Cloud Topology Repair;

[0017] Topology repair is performed using the Poisson reconstruction algorithm, which includes Poisson reconstruction parameter optimization and feature preservation strategies.

[0018] S3, Calculation of normal vector and curvature;

[0019] An adaptive neighborhood PCA algorithm and a GPU parallel acceleration architecture are used to calculate normal vectors and curvature.

[0020] S4. Improved Poisson sampling;

[0021] Poisson sampling optimization is achieved through a sampling function based on multi-feature fusion and a dynamic parameter adjustment strategy.

[0022] S5, Texture curvature segmentation;

[0023] An improved SLIC superpixel algorithm is used to achieve curvature-adaptive segmentation of textures, including energy function construction, parameter setting, and subdivision rule optimization.

[0024] S6, Dynamic Slice Binding;

[0025] Efficient storage, querying, and dynamic binding of slices are achieved through a hierarchical hash table structure and incremental update algorithm, including a three-level query structure design and GPU-coordinated incremental update logic.

[0026] S7, LOD level generation;

[0027] Adaptive generation of hierarchical models is achieved through geometric simplification error control and texture Mipmap optimization;

[0028] S8, dynamically loaded at runtime;

[0029] Efficient scheduling of lightweight models is achieved through view frustum pruning acceleration and intelligent preloading models, including GPU-accelerated visibility judgment and viewpoint prediction-driven resource preloading.

[0030] S9, Boundary blending processing;

[0031] By employing a multi-scale blending strategy and a material-adaptive color difference compensation model, visual discontinuities at the boundaries between different LOD levels and slice blocks are eliminated.

[0032] S10, Output lightweight model;

[0033] Integrate the geometric data, texture data, and LOD level information processed by S1-S9 to output a lightweight 3D model.

[0034] Furthermore, the hybrid acquisition system is equipped with a spatiotemporal synchronization mechanism. This mechanism achieves spatiotemporal synchronization of multiple devices through an FPGA-based hardware synchronization trigger circuit and the PTPV2 protocol. The hybrid acquisition system supports multispectral acquisition. The wavelength of the phase laser scanner is 1550nm±5nm, and the power of the phase laser scanner is adjustable from 10-100mW to adapt to different acquisition scenarios. The timing control accuracy of the FPGA hardware synchronization trigger circuit is not less than 10ns, and the clock deviation of the PTPV2 protocol is not greater than 1us.

[0035] The multispectral acquisition includes the visible light band, UV fluorescence band, and near-infrared band. The visible light band has a wavelength range of 400-700nm and uses a white LED array light source and a 48MP CMOS sensor for conventional texture acquisition. The UV fluorescence band has a wavelength of 365nm and uses a UV-LED light source and a scientific-grade CCD sensor for defect detection. The near-infrared band has a wavelength of 850nm and uses an IR laser light source and an InGaAs sensor for internal structure detection.

[0036] In step S1, coordinate system one is achieved by combining the nine-point calibration method with nonlinear optimization. The objective function of the nonlinear optimization is:

[0037] ,

[0038] In the formula, To calibrate the three-dimensional coordinates of the corner points of the plate, To determine the two-dimensional coordinates of the corner points of the calibration plate in the image, =0.1 is the regularization coefficient, and the calibrated reprojection error is < 0.3 pixels;

[0039] In step S1, an iterative filtering algorithm based on normal consistency is used to denoise the point cloud. The iterative filtering algorithm is set to iterate 3 times. For each point in each iteration, the neighboring points with a radius of 2.5 times the average point spacing are queried through KDTree. The dot product of the normal vector of the neighboring point and the normal vector of the point is calculated. Valid neighboring points with a dot product > 0.8 are selected. When the number of valid neighboring points > 5, the average coordinates of the valid neighboring points are taken to update the point.

[0040] Further, in step S2, the optimization of the Poisson reconstruction parameters is achieved by constructing a relationship model between the octree depth and the reconstruction accuracy. The relationship model is as follows:

[0041] ,

[0042] In the formula, g is the reconstruction error, and d is the octree depth parameter. When scanning cultural relics, the octree depth parameter d=12 and the sampling interval=0.2mm. When conducting industrial inspection, the octree depth parameter d=10 and the sampling interval=0.5mm.

[0043] The feature preservation strategy employs a quadratic error metric for edge features, and the error calculation formula is as follows:

[0044]

[0045] In the formula, Let the area of ​​the triangular facet be... For the normal direction of the surface, For vertices to be simplified, For a set of facets associated with vertices, the feature retention rate of the strategy is ≥96.3%.

[0046] Furthermore, in step S3, the adaptive neighborhood PCA algorithm is implemented by dynamically adjusting the neighborhood radius, and the formula for calculating the neighborhood radius is:

[0047] ,

[0048] In the formula, For local point spacing, For the local radius of curvature, The local radius of curvature is estimated by fitting a local quadratic surface, and then the curvature is calculated based on the quadratic surface. The calculation formula is as follows: In the formula, The Hessian matrix of the quadratic surface. Let be the normal vector of the quadratic surface. The gradient of the quadratic surface. Let be the curvature, and let the local radius of curvature satisfy the following: ;

[0049] The GPU parallel acceleration architecture uses a layered CUDA kernel to implement parallel computation of normal vectors. The computation process of the CUDA kernel includes the following steps:

[0050] S31. Traverse the neighborhood points of each point, accumulate the coordinates of the neighborhood points and calculate the centroid. The centroid is obtained by taking the average value of the number of neighborhood points.

[0051] S32. Use the cuSolver library to perform eigenvalue decomposition on the covariance matrix;

[0052] S33. Select the vector corresponding to the smallest eigenvalue as the normal vector of that point.

[0053] Furthermore, in step S4, the sampling function for multi-feature fusion is implemented through an extended priority function, the calculation formula of which is:

[0054] ,

[0055] In the formula, For point Sampling priority, , , and These are the weight coefficients of each feature term, and they satisfy... + + + =1, Let be the magnitude of the texture gradient vector. For point The curvature value, For point Local density factor, point Edge feature factors;

[0056] In step S4, the dynamic parameter adjustment strategy is implemented by establishing parameter adaptive rules, and the calculation formula for the parameter adaptive rules is as follows:

[0057] ,

[0058] It is the global maximum texture gradient magnitude. Let be the magnitude of the global average texture gradient vector. As the difference in texture variation increases, and The difference widens. These are the texture feature weight coefficients. Automatic boosting increases the weight of texture features in sampling priority, and vice versa; curvature feature weight coefficients. Local density factor weights and edge feature factor weights based on The value is dynamically adjusted proportionally to maintain... + + + =1.

[0059] Furthermore, in step S5, the improved SLIC superpixel algorithm is an energy function that fuses multiple features, expressed as:

[0060] ,

[0061] For total energy loss, These are the pixel coordinates. To preset the number of superpixels, For the first Superpixel clustering For pixels The curvature eigenvector, For the first The curvature feature centers of each cluster, This is the curvature weighting coefficient. For pixels spatial coordinates, For the first The spatial coordinates of each pixel This is a texture gradient penalty term. The texture gradient modulus;

[0062] The segmentation rule optimization is achieved through a dynamic segmentation threshold model, and the calculation formula for the dynamic segmentation threshold is as follows:

[0063] ,

[0064] In the formula, The superpixel subdivision threshold. The average texture gradient magnitude within the region.

[0065] Furthermore, in step S6, the hierarchical hash table structure is a three-level query structure, which is divided into a top level, a middle level, and a bottom level. The top level uses Morton-encoded 64-bit spatial index, the middle level uses RB-Tree to store block bounding boxes, and the bottom level uses a dynamic array to store slice pointers.

[0066] In step S6, the incremental update algorithm is used to dynamically maintain the consistency between the slice data and the hash table. The incremental update algorithm includes the following sub-steps:

[0067] S61. Input the list of blocks to be updated (dirty_list), and iterate through each block in the list;

[0068] S62. Query the old slices (old_slices) currently bound to the block through the hierarchical hash table, and recalculate the new slices (new_slices) of the block.

[0069] S63. Perform a difference analysis on the old slices and the new slices to determine which slices need to be added (added = new_slices - old_slices) and which slices need to be deleted (removed = old_slices - new_slices).

[0070] S64. An 8-thread pool is used to execute the update operation in parallel. The GPU texture update function (update_gpu_texture) is called by multiple threads to load the new slices into the GPU memory. At the same time, the GPU memory release function (free_gpu_memory) is called to reclaim the memory resources occupied by the deleted slices, so as to realize the incremental and efficient update of slice binding.

[0071] Furthermore, in step S7, the geometric simplification error control is implemented based on the quantitative relationship model between QEM error and visual perception. This relationship model is constructed using an exponential function, and its expression is as follows: In the formula, For geometrically simplified perceptibility, Error metrics, model-based LOD hierarchical control strategies are divided into highest level of detail, intermediate level of detail, and basic level of detail, wherein the highest level of detail control... <0.1, corresponding to <0.1, the geometric simplification error is within the range imperceptible to the human eye, and the intermediate level of detail is controlled to be 0.1≤ <0.3, corresponding to 0.1≤ <0.3 indicates a slight, visible geometric simplification error that does not affect core feature recognition; the basic detail level is controlled to be 0.3≤. <0.8, corresponding to 0.3≤ <1.0, allowing for significant errors while maintaining structural integrity;

[0072] In step S7, a neural network-based texture downsampling method is used to adjust the texture accuracy to match the geometric level. The texture downsampling method includes the following steps:

[0073] S71. The encoding module uses the ResNet18 network to extract multi-scale features of texture images, preserving key visual information such as texture edges and color gradients.

[0074] S72. The decoding module maps the encoded features to the target size through the UNet network, generating a downsampled texture that matches the geometric simplification level of the current LOD level.

[0075] S73. Input the original texture image and the target downsampling size. After the encoding module extracts features, the decoding module outputs a downsampling texture that conforms to visual consistency.

[0076] Furthermore, in step S8, the accelerated frustum clipping employs conservative rasterization and HiZ to achieve fast visibility filtering of point cloud blocks. First, the GPU computes the bounding box depth map of each point cloud block in parallel. Then, the association between spatial location and depth information is established, constructing a hierarchical Z-Buffer (HiZ). The visibility of point cloud blocks is tested in parallel. The intelligent preloading model uses a long short-term memory network to construct a view trajectory prediction model.

[0077] In step S9, the multi-scale blending strategy divides the processing range according to the pixel width of the boundary transition region and matches the corresponding blending algorithm. The color difference compensation model is constructed based on the material type and dynamically adjusts the parameters. It achieves adaptive compensation of the original color difference through an exponential function. The model formula is:

[0078] ,

[0079] In the formula, The original color difference value of the boundary region The compensated color difference value, The standard deviation parameter of the Gaussian kernel. The value is dynamically adjusted based on the material type; for metal materials... =1.2px, stone material σ=2.5px, fabric material σ=3px, The distance from the boundary pixel to the transition center

[0080] On the other hand, the present invention also provides a lightweight 3D model system based on point cloud layering and dynamic texture mapping, which includes:

[0081] The data processing and optimization module, through data acquisition, point cloud repair, feature calculation and sampling optimization functions, is used to complete the basic data processing before lightweighting;

[0082] The texture and slice management module constructs a lightweight texture and geometric slice association system through texture segmentation and slice binding;

[0083] The hierarchy generation and loading module integrates LOD hierarchy generation and runtime scheduling functions to balance lightweight efficiency and visual quality.

[0084] The visual consistency optimization and output module integrates boundary processing and model output functions, and is used to output lightweight 3D models;

[0085] The data processing and optimization module includes a hybrid acquisition device and a preprocessing unit. The hybrid acquisition device consists of a phase-type laser scanner and a high dynamic range texture camera, and is used to perform multimodal data synchronous acquisition and spatiotemporal synchronization, coordinate system unification, and point cloud denoising. The preprocessing unit is used to perform point cloud topology repair, normal vector and curvature calculation, and improved Poisson sampling.

[0086] The texture and slice management module includes a texture segmentation unit and a slice binding unit. The texture segmentation unit is used to perform the texture curvature adaptive segmentation, and the slice binding unit is used to perform the dynamic slice binding.

[0087] The hierarchy generation and loading module includes a LOD generation unit and a dynamic loading unit. The LOD generation unit is used to perform the LOD hierarchy generation, and the dynamic loading unit is used to run conservative rasterization and HiZ frustum clipping acceleration, and intelligent preloading of the view trajectory prediction model based on a long short-term memory network.

[0088] The visual consistency optimization and output module includes a boundary processing unit and a model output unit. The boundary processing unit is used for multi-scale blending strategies and material adaptive color difference compensation. The model output unit is used to integrate the geometric, texture and LOD information processed by the above modules and output a lightweight 3D model that meets the requirements of visual consistency and efficient loading.

[0089] Beneficial effects:

[0090] Compared with known public technologies, the technical solution provided by this invention has the following beneficial effects:

[0091] This invention improves storage efficiency, rendering performance, and visual fidelity by establishing a bidirectional coupling mechanism between point clouds and textures. Regarding storage efficiency, it utilizes texture curvature adaptive segmentation and dynamic slicing binding technology to construct an efficient associated storage system for geometry and texture, achieving excellent geometry and texture compression effects and significantly reducing the storage cost of high-precision 3D models. In terms of real-time rendering performance, it significantly improves real-time model scheduling efficiency through runtime dynamic loading strategies combined with efficient visibility filtering acceleration and intelligent resource preloading technology. Regarding feature fidelity, it reduces detail loss in topology repair through edge feature protection strategies, achieves accurate coverage of high-detail areas through multi-feature dynamic sampling, and reduces visual distortion through neural network-driven texture downsampling technology, effectively reducing high-frequency information loss and meeting the core needs of professional fields for high-precision models. Regarding dynamic update efficiency, based on incremental update logic, it avoids the drawback of requiring a full update of model data for dynamic modifications in traditional technologies through precise slice difference analysis and parallel resource scheduling, significantly shortening the time required for local modifications and significantly improving work efficiency in scenarios such as design iteration and model editing. Attached Figure Description

[0092] Figure 1 This is a flowchart of a lightweight 3D model method based on point cloud layering and dynamic texture mapping according to the present invention.

[0093] Figure 2 This is a system diagram of a lightweight 3D model system based on point cloud layering and dynamic texture mapping according to the present invention. Detailed Implementation

[0094] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0095] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0096] The present invention will now be described in further detail with reference to the accompanying drawings:

[0097] Example 1:

[0098] like Figures 1-2 As shown, this invention provides a lightweight 3D model method based on point cloud layering and dynamic texture mapping, including the following steps:

[0099] S1. Data Acquisition and Preprocessing: High-precision acquisition of multimodal data, coordinate system 1, and point cloud denoising provide high-quality data for subsequent topology repair and texture mapping steps.

[0100] Furthermore, multimodal data synchronous acquisition: A hybrid acquisition system consisting of a phase-detection laser scanner and a high dynamic range texture camera is used for multimodal data synchronous acquisition. The hybrid acquisition system is equipped with a spatiotemporal synchronization mechanism, which achieves spatiotemporal synchronization of multiple devices through an FPGA-based hardware synchronization trigger circuit and the PTPV2 protocol. The hybrid acquisition system supports multispectral acquisition. The wavelength of the phase-detection laser scanner is 1550nm±5nm, and the power of the phase-detection laser scanner is adjustable from 10-100mW to adapt to different acquisition scenarios. The timing control accuracy of the FPGA hardware synchronization trigger circuit is no less than 10ns, ensuring no delay in the acquisition actions of the scanner and camera. The clock deviation of the PTPV2 protocol is no greater than 1µs, avoiding point cloud and texture alignment errors caused by clock misalignment. Different spectral acquisition schemes are used for different scenario requirements. Multispectral acquisition includes the visible light band, UV fluorescence band, and near-infrared band. The wavelength range of the visible light band is 400-700nm, using a white LED array light source and a 48MP... CMOS sensors are used for routine texture acquisition, capturing the details of routine textures on the model surface. They are suitable for routine acquisition scenarios such as the patterns on cultural relics and the appearance of mechanical parts. The UV fluorescence band has a wavelength of 365nm and uses a UV-LED light source and a scientific-grade CCD sensor for defect detection. It can detect tiny defects on the model surface by utilizing the fluorescence effect. It is suitable for metal cracks and traces of cultural relic restoration. The near-infrared band has a wavelength of 850nm and uses an IR laser light source and an InGaAs sensor for internal structure detection. Through the IR laser light source and InGaAs sensor, it can penetrate some non-metallic materials to realize the internal structure of the model. It is suitable for detecting internal cavities in wood and delamination of composite materials.

[0101] Furthermore, high-precision calibration and alignment: A coordinate system is achieved through a nine-point calibration method combined with nonlinear optimization. The objective function of the nonlinear optimization is:

[0102] ,

[0103] In the formula, To calibrate the three-dimensional coordinates of the corner points of the plate, Acquired by a scanner, To determine the two-dimensional coordinates of the corner points of the calibration plate in the image, Acquired by camera, =0.1 is the regularization coefficient. =0.1 constrains the orthogonality of the rotation matrix to prevent the optimization results from diverging. For rotation matrix, It is a translation vector. For perspective projection functions, The system uses a 3×3 identity matrix to achieve a unified coordinate system between the scanner and the camera, with a calibrated reprojection error of <0.3 pixels, ensuring accurate alignment of the point cloud and texture in the subsequent process.

[0104] Furthermore, the point cloud denoising algorithm is optimized: an iterative filtering algorithm based on normal consistency is adopted for point cloud denoising. The iterative filtering algorithm is set to iterate 3 times. For each point in each iteration, the neighboring points with a radius of 2.5 times the average point spacing are queried through KDTree. The dot product of the normal vector of the neighboring point and the normal vector of the point is calculated. Valid neighboring points with a dot product > 0.8 are selected. When the number of valid neighboring points > 5, the average coordinates of the valid neighboring points are taken to update the point.

[0105] Furthermore, denoising tests were conducted on the point clouds of two typical models: mechanical parts and cultural relic surfaces. The results are shown in the table below:

[0106] Table 1. Noise Reduction Test Comparison

[0107] Model type Original noise point proportion Residual rate after noise reduction mechanical parts 12.70% 0.30% surface of cultural relics 8.30% 0.10%

[0108] As can be seen from the table above, discrete noise points in the original point cloud were effectively removed, verifying the effectiveness of the point cloud denoising algorithm.

[0109] S2. Point cloud topology repair: Topology repair is performed using the Poisson reconstruction algorithm. Topology repair includes Poisson reconstruction parameter optimization and feature preservation strategies.

[0110] Furthermore, the optimization of Poisson reconstruction parameters is achieved by constructing a relationship model between octree depth and reconstruction accuracy. The relationship model is as follows: In the formula, g is the reconstruction error, and d is the octree depth parameter. When scanning cultural relics, the octree depth parameter d=12 and the sampling interval=0.2mm. When conducting industrial inspection, the octree depth parameter d=10 and the sampling interval=0.5mm.

[0111] Furthermore, the feature preservation strategy employs a quadratic error metric for edge features, with the error calculation formula as follows:

[0112]

[0113] In the formula, Let the area of ​​the triangular facet be... For the normal direction of the surface, For vertices to be simplified, For the set of facets associated with vertices, the feature retention rate of the strategy is ≥96.3%. The feature retention rate is calculated based on the number of edge feature points in the original point cloud model, that is, the proportion of the number of edge feature points retained after topology repair to the number of original edge feature points.

[0114] S3, normal vector and curvature calculation, adopts the adaptive neighborhood PCA algorithm and GPU parallel acceleration architecture to realize normal vector and curvature calculation.

[0115] Furthermore, the adaptive neighborhood PCA algorithm achieves this by dynamically adjusting the neighborhood radius, and the formula for calculating the neighborhood radius is:

[0116] ,

[0117] In the formula, For local point spacing, For the local radius of curvature, The radius of curvature is the neighborhood radius. The local radius of curvature is estimated by fitting a local quadratic surface, and then the curvature is calculated based on the quadratic surface. The calculation formula is as follows: In the formula, The Hessian matrix of the quadratic surface. Let be the normal vector of the quadratic surface. The gradient of the quadratic surface. For curvature, the local radius of curvature satisfies the following relationship with curvature: ;

[0118] Furthermore, the GPU parallel acceleration architecture employs a layered CUDA kernel to achieve parallel computation of normal vectors. The computation process of the CUDA kernel includes the following steps:

[0119] S31. Traverse the neighborhood points of each point, accumulate the coordinates of the neighborhood points and calculate the centroid. The centroid is obtained by taking the average value of the number of neighborhood points.

[0120] S32. Use the cuSolver library to perform eigenvalue decomposition on the covariance matrix;

[0121] S33. Select the vector corresponding to the smallest eigenvalue as the normal vector of that point.

[0122] S4. Improve Poisson sampling by optimizing Poisson sampling through a multi-feature fusion sampling function and a dynamic parameter adjustment strategy.

[0123] Furthermore, for point cloud models with different point counts, the normal vectors and curvatures were calculated using both CPU and the GPU parallel architecture of this method. The time consumption data is shown in the table below:

[0124] Table 2 Comparison of Time Consumption Data

[0125] Points CPU time (ms) GPU time (ms) acceleration ratio 1M 420 12 35× 10M 4,200 85 49×

[0126] As can be seen from the table above, the GPU speedup is up to 49 times, significantly improving computing efficiency.

[0127] S4. Improve Poisson sampling by optimizing Poisson sampling through a multi-feature fusion sampling function and a dynamic parameter adjustment strategy.

[0128] Furthermore, the sampling function for multi-feature fusion is implemented by extending the priority function, and the formula for calculating the priority function is as follows:

[0129] ,

[0130] In the formula, For point Sampling priority, , , and These are the weight coefficients of each feature term, and they satisfy... + + + =1, Let be the magnitude of the texture gradient vector. For point The curvature value, For point Local density factor, point Edge feature factors.

[0131] Furthermore, the dynamic parameter adjustment strategy is implemented by establishing parameter adaptation rules, the calculation formula for which is:

[0132] ,

[0133] It is the global maximum texture gradient magnitude. Let be the magnitude of the global average texture gradient vector. As the difference in texture variation increases, and The difference widens. These are the texture feature weight coefficients. Automatic boosting increases the weight of texture features in sampling priority, and vice versa; curvature feature weight coefficients. Local density factor weights and edge feature factor weights based on The value is dynamically adjusted proportionally to maintain... + + + =1.

[0134] Furthermore, in the digital scene of the roof ridge tiles of the Forbidden City (including two types of areas: dragon relief and flat tile surface), differentiated sampling was achieved through a dynamic parameter adjustment strategy, and the results are shown in the table below:

[0135] Table 3 Sampling Comparison

[0136] Region Type α value α value Sampling density Dragon relief 0.82 0.15 120 points / cm² Flat tile surface 0.55 0.15 35 points / cm²

[0137] As can be seen from the table above, the sampling density of the dragon pattern relief area (high texture detail) is significantly higher than that of the flat tile surface, balancing detail preservation and efficiency.

[0138] S5. Texture curvature segmentation: An improved SLIC superpixel algorithm is used to achieve adaptive curvature segmentation of textures, including energy function construction, parameter setting, and subdivision rule optimization.

[0139] Furthermore, the improved SLIC superpixel algorithm is an energy function that fuses multiple features, expressed as follows:

[0140] ,

[0141] For total energy loss, These are the pixel coordinates. To preset the number of superpixels, For the first Superpixel clustering For pixels The curvature eigenvector, For the first The curvature feature centers of each cluster, This is the curvature weighting coefficient. For pixels spatial coordinates, For the first The spatial coordinates of each pixel This is a texture gradient penalty term. The texture gradient modulus.

[0142] Furthermore, the segmentation rule optimization is achieved through a dynamic segmentation threshold model, and the formula for calculating the dynamic segmentation threshold is:

[0143] ,

[0144] In the formula, The superpixel subdivision threshold. The average texture gradient magnitude within the region.

[0145] Furthermore, in the texture segmentation of two types of industrial parts, gearboxes and turbine blades, the number of slices was optimized through curvature adaptive segmentation, and the results are shown in the table below:

[0146] Table 4 Comparison of Optimized Slices

[0147] Part type Original slice count Optimized number of slices gearbox 1,024 387 （-62%） turbine blades 2,048 612 （-70%）

[0148] As can be seen from the table above, the number of slices is reduced by 62%-70%, significantly reducing storage redundancy.

[0149] S6. Dynamic slice binding: It achieves efficient storage, querying and dynamic binding of slices through a hierarchical hash table structure and incremental update algorithm, including a three-level query structure design and GPU-coordinated incremental update logic.

[0150] Furthermore, the hierarchical hash table structure is a three-level query structure, which is divided into a top level, a middle level, and a bottom level. The top level uses a Morton-encoded 64-bit spatial index, the middle level uses an RB-Tree to store block bounding boxes, and the bottom level uses a dynamic array to store slice pointers.

[0151] Furthermore, the incremental update algorithm is used to dynamically maintain the consistency between the slice data and the hash table. The incremental update algorithm includes the following steps:

[0152] S61. Input the list of blocks to be updated (dirty_list), and iterate through each block in the list;

[0153] S62. Query the old slices (old_slices) currently bound to the block through the hierarchical hash table, and recalculate the new slices (new_slices) of the block.

[0154] S63. Perform a difference analysis on the old slices and the new slices to determine which slices need to be added (added = new_slices - old_slices) and which slices need to be deleted (removed = old_slices - new_slices).

[0155] S64. An 8-thread pool is used to execute the update operation in parallel. The GPU texture update function (update_gpu_texture) is called by multiple threads to load the new slices into the GPU memory. At the same time, the GPU memory release function (free_gpu_memory) is called to reclaim the memory resources occupied by the deleted slices, so as to realize the incremental and efficient update of slice binding.

[0156] Furthermore, for models with different proportions of changed regions, the time consumption of a full update and the incremental update of this invention were compared, and the results are shown in the table below:

[0157] Table 5 Comparison of Update Time

[0158] Change in regional proportion Full update time Incremental update time 5% 320ms 28ms 20% 320ms 65ms

[0159] As can be seen from the table above, incremental updates take only 8.75%-20.31% of the time of full updates, significantly improving the efficiency of partial updates.

[0160] S7, LOD level generation, achieves adaptive generation of level models through geometric simplification error control and texture Mipmap optimization.

[0161] Furthermore, geometrically simplified error control is implemented based on a quantitative relationship model between QEM error and visual perception. This model is constructed using an exponential function, and its expression is as follows: In the formula, For geometrically simplified perceptibility, Error metrics, model-based LOD hierarchical control strategies are divided into highest level of detail, intermediate level of detail, and basic level of detail, with the highest level of detail control... <0.1, corresponding to <0.1 indicates that the geometric simplification error is imperceptible to the human eye, and the medium level of detail control is 0.1≤ <0.3, corresponding to 0.1≤ <0.3 indicates a slight and visible geometric simplification error that does not affect core feature recognition; basic detail level control is 0.3≤ <0.8, corresponding to 0.3≤ <1.0, allowing for significant errors while maintaining structural integrity.

[0162] Furthermore, a neural network-based texture downsampling method is employed to achieve texture precision adjustment that matches the geometric level. The texture downsampling method includes the following steps:

[0163] S71. The encoding module uses the ResNet18 network to extract multi-scale features of texture images, preserving key visual information such as texture edges and color gradients.

[0164] S72. The decoding module maps the encoded features to the target size through the UNet network, generating a downsampled texture that matches the geometric simplification level of the current LOD level.

[0165] S73. Input the original texture image and the target downsampling size. After the encoding module extracts features, the decoding module outputs a downsampling texture that conforms to visual consistency.

[0166] Furthermore, the texture downsampling method of the present invention is compared with the traditional bicubic downsampling method, and the results are shown in the table below:

[0167] Table 6 Sampling Comparison

[0168] method Peak signal-to-noise ratio (dB) Generation time Double three times 36.2 12ms This method 41.7 28ms

[0169] As can be seen from the table above, the peak signal-to-noise ratio of this invention is improved by 5.5dB, resulting in better visual quality. The only difference is a slight increase in generation time, which is within an acceptable range.

[0170] S8 features runtime dynamic loading, achieving efficient scheduling of lightweight models through view frustum pruning acceleration and intelligent preloading models, including GPU-accelerated visibility judgment and viewpoint prediction-driven resource preloading.

[0171] Furthermore, the view frustum clipping acceleration employs conservative rasterization and HiZ to achieve fast visibility filtering of point cloud blocks. First, the GPU computes the bounding box depth map of each point cloud block in parallel. Then, it establishes the correlation between spatial location and depth information, constructing a hierarchical Z-Buffer (HiZ). Parallel testing of point cloud block visibility; intelligent preloading model uses a long short-term memory network to construct a view trajectory prediction model.

[0172] Furthermore, for point cloud scenarios of varying complexity, the time consumption of traditional point-by-point cropping and the HiZ acceleration method of this invention were compared, and the results are shown in the table below:

[0173] Table 7 Comparison of time consumption in different point cloud scenarios

[0174] Scene complexity Traditional cutting (ms) HiZ acceleration (ms) 1 million points 4.2 0.8 10 million points 42 3.5

[0175] As can be seen from the table above, the HiZ method can significantly reduce the time consumption of visibility filtering and is suitable for large-scale point cloud scenarios.

[0176] S9. Boundary blending processing: Through multi-scale blending strategies and a material-adaptive color difference compensation model, visual discontinuities at the boundaries between different LOD levels and slice blocks are eliminated.

[0177] Furthermore, the multi-scale blending strategy divides the processing range according to the pixel width of the boundary transition region and matches the corresponding blending algorithm. The color difference compensation model is constructed based on the material type, dynamically adjusting parameters, and achieving adaptive compensation of the original color difference through an exponential function. The model formula is as follows:

[0178] ,

[0179] In the formula, The original color difference value of the boundary region The compensated color difference value, The standard deviation parameter of the Gaussian kernel. The value is dynamically adjusted based on the material type; for metal materials... =1.2px, stone material σ=2.5px, fabric material σ=3px, This is the distance from the boundary pixel to the transition center.

[0180] Furthermore, the corresponding hybrid algorithms are matched, as detailed in the table below:

[0181] Table 8 Comparison of Hybrid Algorithms

[0182] Mixed Level Scope of processing algorithm Pixel level 0-2px bilateral filtering Block level 2-8px Poisson fusion global level >8px Color migration

[0183] Achieve a seamless transition from pixel-level to global-level mapping.

[0184] S10: Output lightweight model. Integrate the geometric data, texture data, and LOD level information processed by S1-S9 to output a lightweight 3D model.

[0185] Furthermore, the proprietary format output by this invention is compared with mainstream 3D model formats (glTF, USDZ), and the results are shown in the table below:

[0186] Table 9 Comparison of Output Formats

[0187] Format Engine support Geometric compression ratio Texture compression ratio glTF General 10:01 none USDZ apple 3:01 6:01 This method dedicated 18:01 10:01

[0188] As can be seen from the table above, this method significantly outperforms traditional formats in both geometric compression ratio of 18:01 and texture compression ratio of 10:01, and also supports cross-platform adaptive loading.

[0189] The method of this invention was applied to the lightweighting and defect detection of 3D models of aero-engine blades. The key indicators of the traditional method and the method of this invention were compared, and the results are shown in the table below:

[0190] Table 10 Comparison of Detection Indicators

[0191] index Traditional methods This method Model volume 14.8GB 2.3GB (-84%) Defect detection rate 91.20% 98.70% Real-time frame rate 22fps 76fps

[0192] As can be seen from the table above, this method achieves significant improvements in model size, defect detection rate, and real-time frame rate.

[0193] Example 2:

[0194] like Figures 1-2 As shown, Embodiment 2 provides a lightweight 3D model system based on point cloud layering and dynamic texture mapping, which includes:

[0195] The data processing and optimization module, through data acquisition, point cloud repair, feature calculation and sampling optimization functions, is used to complete the basic data processing before lightweighting;

[0196] The texture and slice management module constructs a lightweight texture and geometric slice association system through texture segmentation and slice binding;

[0197] The hierarchy generation and loading module integrates LOD hierarchy generation and runtime scheduling functions to balance lightweight efficiency and visual quality.

[0198] The visual consistency optimization and output module integrates boundary processing and model output functions, and is used to output lightweight 3D models;

[0199] The data processing and optimization module includes a hybrid acquisition device and a preprocessing unit. The hybrid acquisition device consists of a phase laser scanner and a high dynamic range texture camera, which is used to perform multimodal data synchronous acquisition and spatiotemporal synchronization, coordinate system unification, and point cloud denoising. The preprocessing unit is used to perform point cloud topology repair, normal vector and curvature calculation, and improved Poisson sampling.

[0200] The texture and slice management module includes a texture segmentation unit and a slice binding unit. The texture segmentation unit is used to perform texture curvature adaptive segmentation, and the slice binding unit is used to perform dynamic slice binding.

[0201] The hierarchy generation and loading module includes a LOD generation unit and a dynamic loading unit. The LOD generation unit is used to perform LOD hierarchy generation, and the dynamic loading unit is used to accelerate the view frustum clipping of conservative rasterization and HiZ, as well as the intelligent preloading of the view trajectory prediction model based on the long short-term memory network.

[0202] The visual consistency optimization and output module includes a boundary processing unit and a model output unit. The boundary processing unit is used for multi-scale blending strategies and material adaptive color difference compensation, while the model output unit is used to integrate the geometric, texture and LOD information processed by the above modules and output a lightweight 3D model that meets the requirements of visual consistency and efficient loading.

[0203] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightweight 3D model method based on point cloud layering and dynamic texture mapping, characterized in that, Includes the following steps: S1. Data acquisition and preprocessing; A hybrid acquisition system consisting of a phase-detection laser scanner and a high dynamic range texture camera was used for synchronous acquisition of multimodal data. S2, Point Cloud Topology Repair; Topology repair is performed using the Poisson reconstruction algorithm. This topology repair includes Poisson reconstruction parameter optimization and feature preservation strategies. The feature preservation strategy employs a quadratic error metric for edge features, and the error calculation formula is as follows: In the formula, Let the area of ​​the triangular facet be... For the normal direction of the surface, For vertices to be simplified, For a set of facets associated with vertices, the feature retention rate of the strategy is ≥96.3%; S3, Calculation of normal vector and curvature; An adaptive neighborhood PCA algorithm is used to dynamically adjust the neighborhood radius, and a GPU parallel acceleration architecture is used to calculate the normal vector and curvature. S4. Improved Poisson sampling; Poisson sampling optimization is achieved through a multi-feature fusion sampling function and a dynamic parameter adjustment strategy; in step S4, the multi-feature fusion sampling function is implemented by an extended priority function, the calculation formula of which is: In the formula, For point Sampling priority, , , These are the weight coefficients of each feature term, and they satisfy... + + + =1, Let be the magnitude of the texture gradient vector. For point The curvature value, For point Local density factor, point Edge feature factors; In step S4, the dynamic parameter adjustment strategy is implemented by establishing parameter adaptive rules, and the calculation formula for the parameter adaptive rules is as follows: , It is the global maximum texture gradient magnitude. Let be the magnitude of the global average texture gradient vector. As the difference in texture variation increases, and The difference widens. These are the texture feature weight coefficients. Automatic boosting increases the weight of texture features in sampling priority, and vice versa; curvature feature weight coefficients. Local density factor weights and edge feature factor weights based on The value is dynamically adjusted proportionally to maintain... + + + =1; S5, Texture curvature segmentation; An improved SLIC superpixel algorithm is used to achieve curvature-adaptive segmentation of textures, including energy function construction, parameter setting, and subdivision rule optimization. In step S5, the improved SLIC superpixel algorithm is an energy function that fuses multiple features, expressed as follows: ,in, For total energy loss, These are the pixel coordinates. To preset the number of superpixels, For the first Superpixel clustering For pixels The curvature eigenvector, For the first The curvature feature centers of each cluster, This is the curvature weighting coefficient. For pixels spatial coordinates, This is a texture gradient penalty term. The texture gradient modulus; The segmentation rule optimization is achieved through a dynamic segmentation threshold model, and the formula for calculating the dynamic segmentation threshold is as follows: In the formula, The superpixel subdivision threshold. The average texture gradient magnitude within the region; S6, Dynamic Slice Binding; Efficient storage, querying, and dynamic binding of slices are achieved through a hierarchical hash table structure and incremental update algorithm, including a three-level query structure design and GPU-coordinated incremental update logic. S7, LOD level generation; Adaptive generation of hierarchical models is achieved through geometric simplification error control and texture Mipmap optimization; S8, dynamically loaded at runtime; Efficient scheduling of lightweight models is achieved through view frustum pruning acceleration and intelligent preloading models, including GPU-accelerated visibility judgment and viewpoint prediction-driven resource preloading. S9, Boundary blending processing; By employing a multi-scale blending strategy and a material-adaptive color difference compensation model, visual discontinuities at the boundaries between different LOD levels and slice blocks are eliminated. S10, Output lightweight model; Integrate the geometric data, texture data, and LOD level information processed by S1-S9 to output a lightweight 3D model.

2. The method for lightweighting 3D models based on point cloud layering and dynamic texture mapping according to claim 1, characterized in that, In step S1, the hybrid acquisition system is configured with a spatiotemporal synchronization mechanism. The spatiotemporal synchronization mechanism realizes spatiotemporal synchronization of multiple devices through an FPGA-based hardware synchronization trigger circuit and the PTPV2 protocol. The hybrid acquisition system supports multispectral acquisition. The wavelength of the phase laser scanner is 1550nm±5nm. The power adjustable range of the phase laser scanner is 10-100mW to adapt to different acquisition scenarios. The timing control accuracy of the FPGA-based hardware synchronization trigger circuit is not less than 10ns. The clock deviation of the PTPV2 protocol is not greater than 1us. The multispectral acquisition includes the visible light band, UV fluorescence band, and near-infrared band. The visible light band has a wavelength range of 400-700nm and uses a white LED array light source and a 48MP CMOS sensor for conventional texture acquisition. The UV fluorescence band has a wavelength of 365nm and uses a UV-LED light source and a scientific-grade CCD sensor for defect detection. The near-infrared band has a wavelength of 850nm and uses an IR laser light source and an InGaAs sensor for internal structure detection. In step S1, coordinate system one is achieved by combining the nine-point calibration method with nonlinear optimization. The objective function of the nonlinear optimization is: In the formula, To calibrate the three-dimensional coordinates of the corner points of the plate, To determine the two-dimensional coordinates of the corner points of the calibration plate in the image, =0.1 is the regularization coefficient. For rotation matrix, It is a translation vector. For perspective projection functions, It is a 3×3 identity matrix, and the calibrated reprojection error is <0.3 pixels; In step S1, an iterative filtering algorithm based on normal consistency is used to denoise the point cloud. The iterative filtering algorithm is set to iterate 3 times. For each point in each iteration, the neighboring points with a radius of 2.5 times the average point spacing are queried through KDTree. The dot product of the normal vector of the neighboring point and the normal vector of the point is calculated. Valid neighboring points with a dot product > 0.8 are selected. When the number of valid neighboring points > 5, the average coordinates of the valid neighboring points are taken to update the point.

3. The method for lightweighting 3D models based on point cloud layering and dynamic texture mapping according to claim 2, characterized in that, In step S2, the optimization of the Poisson reconstruction parameters is achieved by constructing a relationship model between the octree depth and the reconstruction accuracy. The relationship model is as follows: In the formula, g is the reconstruction error, and d is the octree depth parameter. When scanning cultural relics, the octree depth parameter d=12 and the sampling interval=0.2mm. When conducting industrial inspection, the octree depth parameter d=10 and the sampling interval=0.5mm.

4. The method for lightweighting 3D models based on point cloud layering and dynamic texture mapping according to claim 3, characterized in that, In step S3, the adaptive neighborhood PCA algorithm is implemented by dynamically adjusting the neighborhood radius, and the formula for calculating the neighborhood radius is: In the formula, For local point spacing, For the local radius of curvature, The local radius of curvature is estimated by fitting a local quadratic surface, and then the curvature is calculated based on the quadratic surface. The calculation formula is as follows: In the formula, The Hessian matrix of the quadratic surface. Let be the normal vector of the quadratic surface. The gradient of the quadratic surface. Let be the curvature, and let the local radius of curvature satisfy the following: ; The GPU parallel acceleration architecture uses a layered CUDA kernel to implement parallel computation of normal vectors. The computation process of the CUDA kernel includes the following steps: S31. Traverse the neighborhood points of each point, accumulate the coordinates of the neighborhood points and calculate the centroid. The centroid is obtained by taking the average value of the number of neighborhood points. S32. Use the cuSolver library to perform eigenvalue decomposition on the covariance matrix; S33. Select the vector corresponding to the smallest eigenvalue as the normal vector of that point.

5. A lightweight 3D model method based on point cloud layering and dynamic texture mapping according to claim 4, characterized in that, In step S6, the hierarchical hash table structure is a three-level query structure, which is divided into a top level, a middle level and a bottom level. The top level uses Morton-encoded 64-bit spatial index, the middle level uses RB-Tree to store block bounding boxes, and the bottom level uses a dynamic array to store slice pointers. In step S6, the incremental update algorithm is used to dynamically maintain the consistency between the slice data and the hash table. The incremental update algorithm includes the following sub-steps: S61. Input the list of blocks to be updated, dirty_list, and iterate through each block in the list; S62. Query the old slices currently bound to the block using the hierarchical hash table, and recalculate the new slices of the block. S63. Perform a difference analysis on the old slices and the new slices to determine the slices that need to be added (added = new_slices - old_slices) and the slices that need to be deleted (removed = old_slices - new_slices). S64. An 8-thread pool is used to execute the update operation in parallel. The GPU texture update function update_gpu_texture is called by multiple threads to load the new slices into the GPU memory. At the same time, the GPU memory release function free_gpu_memory is called to reclaim the memory resources occupied by the deleted slices, so as to realize the incremental and efficient update of slice binding.

6. The method for lightweighting 3D models based on point cloud layering and dynamic texture mapping according to claim 5, characterized in that, In step S7, geometric simplification error control is implemented based on a quantitative relationship model between QEM error and visual perception. This relationship model is constructed using an exponential function, and its expression is as follows: In the formula, For geometrically simplified perceptibility, As an error metric, the model-based LOD hierarchical control strategy is divided into highest level of detail, intermediate level of detail, and basic level of detail. The highest level of detail control... <0.1, corresponding to <0.1, the geometric simplification error is within the range imperceptible to the human eye, and the intermediate level of detail is controlled to be 0.1≤ <0.3, corresponding to 0.1≤ <0.3 indicates a slight, visible geometric simplification error that does not affect core feature recognition; the basic detail level is controlled to be 0.3≤. <0.8, corresponding to 0.3≤ <1.0, allowing for significant errors while maintaining structural integrity; In step S7, a neural network-based texture downsampling method is used to adjust the texture accuracy to match the geometric level. The texture downsampling method includes the following steps: S71. The encoding module uses the ResNet18 network to extract multi-scale features of texture images, preserving key visual information such as texture edges and color gradients. S72. The decoding module maps the encoded features to the target size through the UNet network, generating a downsampled texture that is compatible with the geometric simplification level of the current LOD level. S73. Input the original texture image and the target downsampling size. After the encoding module extracts features, the decoding module outputs a downsampling texture that conforms to visual consistency.

7. A lightweight 3D model method based on point cloud layering and dynamic texture mapping according to claim 6, characterized in that, In step S8, the view frustum clipping acceleration uses conservative rasterization and HiZ to achieve fast visibility screening of point cloud blocks. First, the GPU calculates the bounding box depth map of each point cloud block in parallel, then establishes the association between spatial location and depth information, constructs a hierarchical Z-Buffer (HiZ), and tests the visibility of point cloud blocks in parallel. The intelligent preloading model uses a long short-term memory network to construct a view trajectory prediction model. In step S9, the multi-scale blending strategy divides the processing range according to the pixel width of the boundary transition region and matches the corresponding blending algorithm. The color difference compensation model is constructed based on the material type and dynamically adjusts the parameters. It achieves adaptive compensation of the original color difference through an exponential function. The model formula is: In the formula, The original color difference value of the boundary region. The compensated color difference value, The standard deviation parameter of the Gaussian kernel. The value is dynamically adjusted based on the material type; for metal materials... =1.2px, stone material σ=2.5px, fabric material σ=3px, This is the distance from the boundary pixel to the transition center.

8. A lightweight 3D model system based on point cloud layering and dynamic texture mapping, comprising the lightweight 3D model method based on point cloud layering and dynamic texture mapping as described in any one of claims 1-7, characterized in that, It includes: The data processing and optimization module, through data acquisition, point cloud repair, feature calculation and sampling optimization functions, is used to complete the basic data processing before lightweighting; The texture and slice management module constructs a lightweight texture and geometric slice association system through texture segmentation and slice binding; The hierarchy generation and loading module integrates LOD hierarchy generation and runtime scheduling functions to balance lightweight efficiency and visual quality. The visual consistency optimization and output module integrates boundary processing and model output functions, and is used to output lightweight 3D models; The data processing and optimization module includes a hybrid acquisition device and a preprocessing unit. The hybrid acquisition device consists of a phase-type laser scanner and a high dynamic range texture camera, and is used to perform multimodal data synchronous acquisition and spatiotemporal synchronization, coordinate system unification, and point cloud denoising. The preprocessing unit is used to perform point cloud topology repair, normal vector and curvature calculation, and improved Poisson sampling. The texture and slice management module includes a texture segmentation unit and a slice binding unit. The texture segmentation unit is used to perform the texture curvature adaptive segmentation, and the slice binding unit is used to perform the dynamic slice binding. The hierarchy generation and loading module includes a LOD generation unit and a dynamic loading unit. The LOD generation unit is used to perform the LOD hierarchy generation, and the dynamic loading unit is used to run conservative rasterization and HiZ frustum clipping acceleration, and intelligent preloading of the view trajectory prediction model based on a long short-term memory network. The visual consistency optimization and output module includes a boundary processing unit and a model output unit. The boundary processing unit is used for multi-scale blending strategies and material adaptive color difference compensation. The model output unit is used to integrate the geometric, texture and LOD information processed by the above modules and output a lightweight 3D model that meets the requirements of visual consistency and efficient loading.

Images