Vehicle depot quasi-real-time 3d modeling method based on Gaussian Splitters
Through the Gaussian Splattings-based method, vehicle depot data is collected and optimized in real time, and three-dimensional models are generated and updated. The problem of inconsistent 3D modeling and actual situations is solved, the accuracy and timeliness of the model are improved, and the efficient management of vehicle depots is supported.
Patent Information
- Application Number
- CN202510633079.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-16
- Publication Date
- 2025-08-19
AI Technical Summary
The existing 3D modeling methods of vehicle depots cannot track the updates and transformations of equipment and facilities in real time, resulting in differences between the models and actual situations, affecting the accuracy of operational and management decisions, especially in scenarios where frequent updates or monitoring are performed.
The Gaussian Splattings-based method is adopted to collect multi-source data in real time, perform data preprocessing, feature extraction and optimization, and use the Gaussian Splatting algorithm to generate a three-dimensional surface, and combine real-time update and optimization mechanisms to dynamically update the three-dimensional model.
Real-time accuracy and timeliness of the three-dimensional model of the vehicle depot are achieved, which can better reflect actual changes, improve operation and management support, and enhance the accuracy and detailed performance of the model.
Smart Images
Figure CN120510294A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of 3D modeling, and in particular to a quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings. Background Art
[0002] A rolling stock depot is a crucial component of railway operations and urban rail transit systems (such as subways and light rail transit). It is primarily responsible for the operation, maintenance, and overhaul of train vehicles (excluding locomotives). 3D modeling of rolling stock depots is a new technology being applied to depot management with the advancement of digital technology. It utilizes 3D modeling technology to accurately digitize various equipment and assets within the depot, enabling detailed representation of everything from large maintenance equipment to tiny components within the depot.
[0003] Through 3D modeling, depot managers can intuitively view vehicle parking, track layout, and equipment distribution without relying on traditional 2D drawings or on-site inspections. This significantly improves management efficiency and reduces misunderstandings and communication costs. Furthermore, technicians can clearly view the structure and details of vehicles and equipment, assisting with maintenance and repair work, thereby improving maintenance efficiency and reducing failures caused by misoperation or omissions. Furthermore, 3D models enable various simulations and analyses, such as vehicle scheduling and emergency evacuation simulations. These results provide managers with a basis for scientific decision-making, improving the rationality and accuracy of their decisions.
[0004] However, in actual application, since the equipment and facilities of the vehicle depot are often updated and modified, the existing 3D modeling often cannot keep up with the pace of actual changes. This may lead to certain differences between the model and the actual status of the vehicle depot. Especially for those scenarios that require frequent updates or monitoring, such as vehicle depots and transportation hubs, this lag makes it impossible for relevant personnel to understand the actual situation in a timely manner, thereby affecting the accuracy of the vehicle depot's operations and management decisions. Summary of the Invention
[0005] The purpose of the present invention is to provide a quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings to solve the problems raised in the above background technology.
[0006] To achieve the above object, the present invention provides the following technical solution: a quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings, comprising the following steps:
[0007] S1. Data acquisition and preprocessing: Sensors are used to collect multi-source data from the depot in real time, including point clouds (representing the shape and position of objects in three-dimensional space) and images (providing visual information of the depot). The collected data is then preprocessed to ensure accuracy and consistency.
[0008] S2. Feature extraction and point cloud optimization: Extract key features from preprocessed data and optimize point cloud data to reduce data redundancy and improve processing efficiency;
[0009] S3, Gaussian Splatting processing: Convert the optimized point cloud data into a Gaussian distribution. Each point is assigned a three-dimensional Gaussian distribution. The Gaussian Splatting algorithm is then applied to "splash" the point cloud data into three-dimensional space. By calculating the interaction between each Gaussian distribution and its neighboring points, a three-dimensional surface is generated.
[0010] S4. Surface reconstruction and mesh generation: Use the sputtered point cloud data to construct a 3D surface model of the vehicle depot and smooth the generated mesh to improve the model's visual quality and accuracy.
[0011] S5. Real-time update and optimization: Dynamically update the collected data according to the real-time changes in the vehicle depot (such as vehicle movement, addition of new equipment, etc.), and process the updated data to update the 3D model in real time. At the same time, optimize the model to maintain the real-time and accuracy of the model.
[0012] Furthermore, in step S1, data is collected from multiple angles and positions by using mobile sensors or multiple fixed sensors to obtain global information of the vehicle segment, and the sensors collect data in real time at a certain frequency to ensure the timeliness and accuracy of the data. The sensor types include but are not limited to laser radar (LiDAR), camera, inertial navigation system (INS), and global positioning system (GPS), as follows:
[0013] The LiDAR measures distance by emitting laser beams and receiving reflected signals, thereby generating three-dimensional point cloud data of the vehicle segment. This data can accurately represent the position and shape of objects such as vehicles, tracks, and buildings.
[0014] The camera captures image data of the vehicle segment to provide visual information. These images are used to extract features such as the vehicle's outline, color, and texture, as well as for subsequent texture mapping and detail enhancement.
[0015] The inertial navigation system and the global positioning system provide positioning and navigation information of the vehicle segment, and are used to determine the position of the collected data in the global coordinate system.
[0016] Furthermore, in step S1, the raw data is preprocessed using a point cloud processing library (PCL, an open source point cloud processing library that provides a wealth of algorithms and tools for point cloud data preprocessing, feature extraction, segmentation, registration, etc.) or an image processing library (OpenCV, an open source computer vision library that can be used for image data preprocessing, feature extraction, image matching, etc.), including the following steps:
[0017] Denoising: Use filtering algorithms (such as Gaussian filtering, mean filtering, etc.) to remove noise points in the data caused by sensor errors, environmental interference, etc., to improve the accuracy and reliability of the data;
[0018] Filtering: Smoothing data through methods such as Laplace smoothing and median filtering can reduce errors and fluctuations, help extract clearer features, and improve the efficiency of subsequent processing;
[0019] Registration: Using ICP (Iterative Closest Point) algorithm, NDT (Normal Distribution Transformation) algorithm, etc., the position and posture of the point cloud are automatically adjusted, and data collected from different sensors or at different time points are aligned to the same coordinate system to ensure data consistency and comparability;
[0020] Time synchronization: During the data collection process, a timestamp is added to each data point. These timestamps record the specific time point when the data was collected. By comparing the timestamps of different data points, time synchronization between data is achieved, ensuring the consistency of all data on the timeline, so as to analyze the dynamic changes of the vehicle segment, such as the movement trajectory of the vehicle.
[0021] Furthermore, in step S2, the key features include the outline of the vehicle, the edge of the track, the structure of the building, etc., and the feature extraction method includes:
[0022] Geometric feature-based methods: use the geometric features of point cloud data (such as normals, curvature, etc.) to extract features. The specific operation is to identify the significant change areas in the point cloud based on the local neighborhood information of the point cloud data, thereby extracting key features;
[0023] Machine learning-based methods: Use machine learning algorithms (such as support vector machines (SVMs) and random forests (RFs)) to classify and extract features from point cloud data. Specific operations include pre-training models to learn feature representations in the data and using these features for classification and recognition in actual applications.
[0024] Combined with deep learning methods: Use deep learning networks (such as convolutional neural networks (CNN), point cloud neural networks (PointNet), etc.) to extract features from point cloud data. Specific operations: Use deep learning networks to automatically learn hierarchical features in the data, thereby extracting higher-level feature representations for subsequent 3D modeling and recognition tasks.
[0025] Furthermore, in step S2, point cloud optimization reduces the amount of point cloud data and reduces data complexity by downsampling or key point extraction, while retaining the overall shape and key features of the point cloud.
[0026] Furthermore, in step S3, the three-dimensional Gaussian distribution is defined by the following parameters:
[0027] Mean: represents the center point of the Gaussian distribution, corresponding to the three-dimensional coordinates of each point in the point cloud;
[0028] Variance: describes the extent to which the Gaussian distribution spreads in all directions. Its size can be adjusted according to the density and characteristics of the point cloud to ensure that the generated surface is both smooth and retains key details;
[0029] Rotation parameter: used to define the direction of the Gaussian distribution in three-dimensional space;
[0030] By assigning a three-dimensional Gaussian distribution with the above parameters to each point in the point cloud, the discrete point cloud data can be converted into a continuous Gaussian field. These parameters allow the Gaussian distribution to have different shapes and extensions in different directions, thereby more flexibly adapting to the complex morphology of the point cloud.
[0031] Furthermore, in step S3, the application of the Gaussian Splatting algorithm includes the following process operations:
[0032] Splatting: Spread each Gaussian distribution into three-dimensional space. Specifically, the Gaussian distribution is discretized within its extended range and the value of each discrete point (i.e., the density of the Gaussian distribution) is stored in a three-dimensional data structure (such as a voxel grid).
[0033] Interaction calculation: By evaluating the degree of overlap of different Gaussian distributions at the same spatial location, the interaction between each Gaussian distribution and its neighboring points is calculated. The higher the degree of overlap, the more significant the surface feature at that location.
[0034] Surface generation: Based on the interaction results of Gaussian distribution, interpolation or reconstruction operations are performed on the three-dimensional data structure to generate a smooth surface, that is, a three-dimensional surface.
[0035] Furthermore, the step S4 specifically includes the following sub-steps:
[0036] S41. Triangulation: Use triangulation algorithms (such as Delaunay triangulation, Ball-Pivoting algorithm, Poisson surface reconstruction, etc.) to connect discrete points in the point cloud data into a triangular mesh;
[0037] S42. Mesh generation: Based on the triangulation results, determine the mesh resolution, topology, and geometry, and generate a preliminary triangular mesh.
[0038] S43, Mesh Smoothing: Use smoothing algorithms (such as Laplace smoothing, mean smoothing, etc.) to smooth the generated mesh, adjust the positions of mesh vertices, reduce fluctuations and unevenness on the mesh surface, and improve the visual effect and accuracy of the model;
[0039] S44. Mesh optimization: Based on smoothing, the mesh is optimized, including deleting redundant vertices, merging adjacent triangles, adjusting the mesh's topological structure, etc.
[0040] Furthermore, in step S5, the model optimization includes mesh fineness adjustment and texture update, as follows:
[0041] Grid fineness adjustment: Dynamically adjust the grid fineness according to the real-time changes of the vehicle depot;
[0042] Texture Updates: As the vehicle depot changes, texture information is updated in real time, including updates to vehicle appearance and texture mapping for new equipment, keeping the model synchronized with the real world.
[0043] Furthermore, the grid fineness adjustment strategy is as follows: in areas where vehicles frequently move or new equipment are added, the grid density is increased to improve the accuracy of the model; in relatively stable areas, the grid density is appropriately reduced to improve rendering efficiency.
[0044] The present invention provides a quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings, which has the following beneficial effects:
[0045] By collecting multi-source data of the vehicle depot in real time, the present invention can more accurately capture and represent the real-time changes of the vehicle depot, generate a realistic and detailed three-dimensional model, and dynamically update the three-dimensional model in conjunction with the real-time update and optimization mechanism, ensuring that the model is always consistent with the actual situation, improving the timeliness and accuracy of the model. It is particularly suitable for scenarios that require frequent updates or monitoring, and can provide more powerful support for the operation and management of the vehicle depot. In addition, compared with existing modeling methods, the accuracy of this method is significantly improved. By extracting key features from the original data and finely optimizing the point cloud data, the accuracy and detail of the three-dimensional model are effectively improved, thereby helping to more accurately reflect the actual structure and status of the vehicle depot. Combined with the application of the Gaussian Splatting algorithm, the optimized point cloud data is converted into the form of a Gaussian distribution. This processing method not only helps to generate smoother and more continuous three-dimensional surfaces, but also improves the processing efficiency of the algorithm, making the modeling process more efficient. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 A schematic flow chart of the steps of a quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to the present invention;
[0047] Figure 2 This is a schematic diagram of step S4 of a quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings of the present invention. DETAILED DESCRIPTION
[0048] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.
[0049] like Figure 1-Figure 2 As shown, a quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings includes the following steps:
[0050] S1. Data acquisition and preprocessing: Use sensors such as lidar, cameras, inertial navigation systems, and global positioning systems to collect multi-source data from the vehicle segment in real time, including point clouds (representing the shape and position of objects in three-dimensional space), images (providing visual information of the vehicle segment), etc. In actual operation, use mobile sensors or multiple fixed sensors to collect data from multiple angles and positions at a certain frequency in real time to ensure that the global information of the vehicle segment, as well as the timeliness and accuracy of the data, are obtained. Then, the collected data is preprocessed to ensure the accuracy and consistency of the data. Specifically, the point cloud processing library (PCL, an open source point cloud processing library, provides a wealth of algorithms and tools for point cloud data preprocessing, feature extraction, segmentation, alignment, etc.) or image processing library (OpenCV, an open source computer vision library that can be used for image data preprocessing, feature extraction, image matching, etc.) is used to perform the following operations on the raw data:
[0051] Denoising: Use filtering algorithms (such as Gaussian filtering, mean filtering, etc.) to remove noise points in the data caused by sensor errors, environmental interference, etc., to improve the accuracy and reliability of the data;
[0052] Filtering: Smoothing data through methods such as Laplace smoothing and median filtering can reduce errors and fluctuations, help extract clearer features, and improve the efficiency of subsequent processing;
[0053] Registration: Using ICP (Iterative Closest Point) algorithm, NDT (Normal Distribution Transformation) algorithm, etc., the position and posture of the point cloud are automatically adjusted, and data collected from different sensors or at different time points are aligned to the same coordinate system to ensure data consistency and comparability;
[0054] Time synchronization: During the data collection process, a timestamp is added to each data point. These timestamps record the specific time point when the data was collected. By comparing the timestamps of different data points, time synchronization between data is achieved, ensuring the consistency of all data on the timeline, so as to analyze the dynamic changes of the vehicle segment, such as the movement trajectory of the vehicle.
[0055] S2. Feature extraction and point cloud optimization: Extract key features from the preprocessed data, including the outline of the vehicle, the edge of the track, the structure of the building, etc., and optimize the point cloud data. By downsampling or extracting key points, the amount of point cloud data can be reduced and the data complexity can be reduced. At the same time, the overall shape and key features of the point cloud are retained to reduce data redundancy and improve processing efficiency.
[0056] In this embodiment, this step may adopt the following feature extraction method:
[0057] Geometric feature-based methods: use the geometric features of point cloud data (such as normals, curvature, etc.) to extract features. The specific operation is to identify the significant change areas in the point cloud based on the local neighborhood information of the point cloud data, thereby extracting key features;
[0058] Machine learning-based methods: Use machine learning algorithms (such as support vector machines (SVMs) and random forests (RFs)) to classify and extract features from point cloud data. Specific operations include pre-training models to learn feature representations in the data and using these features for classification and recognition in actual applications.
[0059] Combined with deep learning methods: Use deep learning networks (such as convolutional neural networks (CNN), point cloud neural networks (PointNet), etc.) to extract features from point cloud data. Specific operations: Use deep learning networks to automatically learn hierarchical features in the data, thereby extracting higher-level feature representations for subsequent 3D modeling and recognition tasks.
[0060] S3, Gaussian Splatting processing: The optimized point cloud data is converted into a Gaussian distribution. Each point is assigned a three-dimensional Gaussian distribution, and the Gaussian Splatting algorithm is applied to "sputter" the point cloud data into three-dimensional space. By calculating the interaction between each Gaussian distribution and its neighboring points, a three-dimensional surface is generated.
[0061] In this embodiment, the three-dimensional Gaussian distribution is defined by the following parameters:
[0062] Mean: represents the center point of the Gaussian distribution, corresponding to the three-dimensional coordinates of each point in the point cloud;
[0063] Variance: describes the extent to which the Gaussian distribution spreads in all directions. Its size can be adjusted according to the density and characteristics of the point cloud to ensure that the generated surface is both smooth and retains key details;
[0064] Rotation parameter: used to define the direction of the Gaussian distribution in three-dimensional space;
[0065] By assigning a three-dimensional Gaussian distribution with the above parameters to each point in the point cloud, the discrete point cloud data can be converted into a continuous Gaussian field. These parameters allow the Gaussian distribution to have different shapes and extensions in different directions, thereby more flexibly adapting to the complex morphology of the point cloud.
[0066] In this embodiment, the Gaussian Splatting algorithm application includes the following process operations:
[0067] Splatting: Spread each Gaussian distribution into three-dimensional space. Specifically, the Gaussian distribution is discretized within its extended range and the value of each discrete point (i.e., the density of the Gaussian distribution) is stored in a three-dimensional data structure (such as a voxel grid).
[0068] Interaction calculation: By evaluating the degree of overlap of different Gaussian distributions at the same spatial location, the interaction between each Gaussian distribution and its neighboring points is calculated. The higher the degree of overlap, the more significant the surface feature at that location.
[0069] Surface generation: Based on the interaction results of Gaussian distribution, interpolation or reconstruction operations are performed on the three-dimensional data structure to generate a smooth surface, that is, a three-dimensional surface.
[0070] S4. Surface reconstruction and mesh generation: Use the point cloud data after sputtering to construct a 3D surface model of the vehicle depot, and smooth the generated mesh to improve the visual effect and accuracy of the model.
[0071] S41. Triangulation: Use triangulation algorithms (such as Delaunay triangulation, Ball-Pivoting algorithm, Poisson surface reconstruction, etc.) to connect discrete points in the point cloud data into a triangular mesh;
[0072] S42. Mesh generation: Based on the triangulation results, determine the mesh resolution, topology, and geometry, and generate a preliminary triangular mesh.
[0073] S43, Mesh Smoothing: Use smoothing algorithms (such as Laplace smoothing, mean smoothing, etc.) to smooth the generated mesh, adjust the positions of mesh vertices, reduce fluctuations and unevenness on the mesh surface, and improve the visual effect and accuracy of the model;
[0074] S44. Mesh optimization: Based on smoothing, the mesh is optimized, including deleting redundant vertices, merging adjacent triangles, adjusting the mesh's topological structure, etc.
[0075] S5. Real-time update and optimization: Dynamically update the collected data based on real-time changes in the vehicle depot (such as vehicle movement, addition of new equipment, etc.), and process the updated data to update the 3D model in real time. At the same time, the model is optimized, including mesh fineness adjustment and texture update.
[0076] Mesh fineness adjustment: Dynamically adjust the mesh fineness based on real-time changes in the vehicle depot; in areas where vehicles move frequently or new equipment is added, increase the mesh density to improve model accuracy; in relatively stable areas, appropriately reduce the mesh density to improve rendering efficiency.
[0077] Texture Updates: As the vehicle depot changes, texture information is updated in real time, including updates to vehicle appearance and texture mapping for new equipment, keeping the model synchronized with the real world.
[0078] The embodiments of the present invention are presented for purposes of illustration and description and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those skilled in the art. The embodiments are chosen and described in order to better illustrate the principles of the invention and its practical application and to enable those skilled in the art to understand the invention and design various embodiments with various modifications as suited for specific applications.
Claims
1. A quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings, characterized by: The following steps are involved: S1. Data collection and preprocessing: Use sensors to collect multi-source data from the vehicle depot in real time and preprocess the collected data; S2, Feature Extraction and Point Cloud Optimization: Extract key features from the preprocessed data and optimize the point cloud data; S3, Gaussian Splatting processing: Convert the optimized point cloud data into Gaussian distribution and apply Gaussian Splatting algorithm to generate a three-dimensional surface; S4, surface reconstruction and mesh generation: Use the point cloud data after sputtering to construct a 3D surface model of the vehicle depot and smooth the generated mesh; S5. Real-time update and optimization: Dynamically update the collected data according to the real-time changes in the vehicle depot, and process the updated data to update the 3D model in real time and optimize the model at the same time.
2. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: In step S1, data is collected from multiple angles and positions by using mobile sensors or multiple fixed sensors to obtain global information of the vehicle segment, and the sensors collect data in real time at a certain frequency. The sensor types include lidar, camera, inertial navigation system, and global positioning system.
3. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: In step S1, the original data is preprocessed using a point cloud processing library or an image processing library. The following steps are involved: Denoising: Use filtering algorithms to remove noise points in the data; Filtering: Smoothing the data through Laplace smoothing and median filtering to reduce errors and fluctuations; Registration: Using ICP and NDT algorithms, the position and posture of the point cloud are automatically adjusted to align data collected from different sensors or at different time points into the same coordinate system. Time synchronization: During the data collection process, a timestamp is added to each data point, and time synchronization between data is achieved by comparing the timestamps of different data points.
4. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: In step S2, the key features include the outline of the vehicle, the edge of the track, and the structure of the building, and the feature extraction method includes: Geometric feature-based method: uses the geometric characteristics of point cloud data to extract features. The specific operation is to identify the significant change areas in the point cloud based on the local neighborhood information of the point cloud data, thereby extracting key features. Machine learning-based methods: Use machine learning algorithms to classify and extract features from point cloud data. Specific operations include pre-training models to learn feature representations in the data and using these features for classification and recognition in actual applications. Combined with deep learning methods: Use deep learning networks to extract features from point cloud data. Specific operations: Use deep learning networks to automatically learn hierarchical features in the data, thereby extracting higher-level feature representations for subsequent 3D modeling and recognition tasks.
5. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: In step S2, point cloud optimization reduces the amount of point cloud data and reduces data complexity by downsampling or key point extraction, while retaining the overall shape and key features of the point cloud.
6. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: In step S3, the three-dimensional Gaussian distribution is defined by the following parameters: Mean: represents the center point of the Gaussian distribution, corresponding to the three-dimensional coordinates of each point in the point cloud; Variance: describes the extent to which the Gaussian distribution spreads in all directions; Rotation parameter: used to define the direction of the Gaussian distribution in three-dimensional space.
7. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: In step S3, the application of the Gaussian Splatting algorithm includes the following process operations: Sputtering: Sputter each Gaussian distribution into three-dimensional space. Specifically, the Gaussian distribution is discretized within its extended range and the value of each discrete point is stored in a three-dimensional data structure. Interaction calculation: By evaluating the degree of overlap of different Gaussian distributions at the same spatial position, the interaction between each Gaussian distribution and its neighboring points is calculated; Surface generation: Based on the interaction results of Gaussian distribution, interpolation or reconstruction operations are performed on the three-dimensional data structure to generate a smooth surface, that is, a three-dimensional surface.
8. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: The step S4 specifically includes the following sub-steps: S41. Triangulation: Use triangulation algorithm to connect discrete points in point cloud data into triangular meshes; S42. Mesh generation: Based on the triangulation results, determine the mesh resolution, topology, and geometry, and generate a preliminary triangular mesh. S43, Mesh Smoothing: Use smoothing algorithms to smooth the generated mesh, adjust the positions of mesh vertices, and reduce fluctuations and unevenness on the mesh surface; S44. Mesh optimization: Based on smoothing, the mesh is optimized, including deleting redundant vertices, merging adjacent triangles, and adjusting the topological structure of the mesh.
9. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 1, characterized in that: In step S5, model optimization includes mesh refinement adjustment and texture update, as follows: Grid fineness adjustment: Dynamically adjust the grid fineness according to the real-time changes of the vehicle depot; Texture Updates: As the vehicle segment changes, texture information is updated in real time, including updates to vehicle appearance and texture mapping for new equipment, keeping the model synchronized with the real world.
10. The quasi-real-time 3D modeling method for a vehicle depot based on Gaussian Splattings according to claim 9, characterized in that: The grid fineness adjustment strategy is as follows: in areas where vehicles frequently move or new equipment is added, the grid density is increased; in relatively stable areas, the grid density is reduced.