Real-scene three-dimensional scene construction method

Abstract

The invention relates to the field of spatial information processing, and discloses a live-action three-dimensional scene construction method, which comprises the following steps: S1, collecting multi-source heterogeneous data of a scene to be constructed, including image data, laser radar point cloud data and geographic information data; according to the method, high-precision rigid registration between the laser radar point cloud data and the image and geographic information data is realized by introducing an improved iterative nearest point algorithm, and on the basis of a traditional ICP algorithm, a dynamic weight adjustment mechanism is adopted, a weighting matrix is introduced into a target function for minimizing registration residual errors, so that the registration precision of the laser radar point cloud data is improved. According to the method, the registration contribution degree can be adaptively adjusted according to the quality, density and spatial distribution characteristics of different data sources, the spatial fusion precision between heterogeneous data is effectively improved, and the risk of model offset or distortion caused by sensor errors is reduced.

Description

Technical Field

[0001] The present invention relates to the field of spatial information processing technology, and in particular to a method for constructing a real-life three-dimensional scene. Background Art

[0002] The method of real-life 3D scene construction is used to restore real-world scenes into high-precision 3D models in computers through technical means. Its function is to collect real-world images, lidar point clouds, geographic information and other data, and after processing, construct a 3D digital model with real appearance, spatial structure and semantic information. This model can not only be viewed, but also measured, annotated, interacted and analyzed. It is widely used in smart city construction, virtual reality, autonomous driving simulation, security monitoring, digital cultural heritage protection and other fields.

[0003] According to Chinese patent publication number CN117788688A, a method for constructing a real-life 3D scene is disclosed. The basic description is as follows: first configure the initial layer of the 3D scene, then configure the initial tool of the 3D scene, then build a new scene, and then configure the base map, terrain layer, business layer, custom data and tool of the new scene. After the configuration of the new scene is completed, use the scene preview function to check whether the new scene meets the requirements. If the new scene meets the requirements, share the new scene with other users through the scene publishing tool. This method is simple and easy to use, the method is clear and concise, and does not require any back-end permission support. Any user can build an online scene based on the system preset data and the user's own customized data, and the constructed scene data is completely isolated from user permissions. The data is secure and can be easily and quickly shared based on the system preset data and user customized data.

[0004] In practice, there are some problems:

[0005] 1. Traditional methods for registering LiDAR point clouds with imagery and geographic information data typically rely on the traditional iterative closest point algorithm, which primarily uses the geometric features of the point cloud to achieve registration. However, the ICP algorithm is susceptible to factors such as the quality, density, and noise of different data sources, resulting in insufficient registration accuracy. This is particularly problematic in complex environments, where it struggles to effectively integrate low-quality or irregular data.

[0006] 2. Traditional 3D modeling methods primarily focus on geometric reconstruction, neglecting the semantic understanding of objects within a scene. This results in 3D models lacking sufficient intelligence and analyzability in applications such as urban modeling, automatic annotation, and scene recognition, making them ineffective in supporting subsequent intelligent analysis and decision-making.

[0007] 3. Traditional 3D modeling methods usually load the texture data of the entire scene at once during texture mapping, which causes the system to be overloaded. Especially when switching perspectives, frequent texture loading will cause obvious delays and flickering, affecting the user experience.

[0008] 4. Traditional 3D modeling methods are typically based on single-shot data collection. They cannot process data changes over a continuous time period in real time, nor can they adapt to the addition, movement, or disappearance of objects in dynamic scenes. This modeling approach cannot cope with the demands of time series changes and cannot provide real-time scene updates. Summary of the Invention

[0009] In response to the shortcomings of the existing technology, the present invention provides a method for constructing real-life three-dimensional scenes, which solves the problems of low registration accuracy of lidar point clouds with other data sources, insufficient semantic expression capabilities of traditional three-dimensional modeling systems, and low efficiency of texture mapping and loading, which lead to rendering delays.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for constructing a real-life three-dimensional scene, comprising the following steps:

[0011] S1. Collect multi-source heterogeneous data of the scene to be constructed, including image data, lidar point cloud data and geographic information data;

[0012] S2, spatially register and fuse multi-source data to form a multidimensional data set in a unified coordinate system;

[0013] S3: Build an initial 3D model based on the fused dataset and perform semantic segmentation and classification on the model;

[0014] S4, using dynamic texture mapping technology to perform multi-scale real-time rendering of 3D models;

[0015] S5. The constructed three-dimensional scene supports multi-terminal access, real-time interaction and intelligent analysis applications.

[0016] Preferably, the spatial registration in step S2 includes:

[0017] The point cloud data is rigidly registered based on the improved iterative closest point algorithm, in which a weighting mechanism is introduced to optimize the fusion accuracy between different data sources;

[0018] The registration process is achieved by minimizing the weighted residual square sum objective function, which is formulated as follows:

[0019]

[0020] Among them, wi is the dynamic weight, pi and qi are the corresponding point pairs in the source point cloud and the target point cloud respectively.

[0021] R is the rotation matrix and t is the translation vector.

[0022] Preferably, the semantic segmentation in S3 adopts a convolutional neural network including an adaptive attention mechanism, specifically including:

[0023] The encoder uses multi-scale convolution to extract spatial features, and the decoder introduces channel and spatial attention modules to improve boundary recognition.

[0024] Cross-layer feature fusion is achieved through residual connections to enhance the segmentation accuracy of small targets and complex backgrounds.

[0025] Preferably, the dynamic texture mapping in S4 includes:

[0026] High-frequency areas use a level-of-detail control mechanism to adjust texture resolution in real time based on viewing distance;

[0027] Texture caching and heat prediction mechanisms are introduced in low-frequency areas to improve rendering efficiency.

[0028] Preferably, the texture prediction mechanism adopts a sliding window algorithm based on the user's viewing history to predict the viewing angle change trend and preload the target area texture, specifically including:

[0029] Record the user perspective vector sequence {ut-N+1,.....,ut} for the past N frames;

[0030] Use the linear prediction model to estimate the viewing direction of the next frame. The formula is as follows:

[0031]

[0032] Among them, α∈(0,1).

[0033] Preferably, an incremental modeling mechanism based on time series changes is introduced during the construction process, including:

[0034] Compare the data collected at different time periods in time to detect the addition, movement and deletion of objects;

[0035] The 3D model is updated using a differential modeling module, and the update is based on a sparse difference extraction strategy using hash matching.

[0036] Preferably, the multi-source heterogeneous data collection adopts an asynchronous collection scheduling mechanism, including:

[0037] Adopt energy consumption priority scheduling strategy based on the remaining power of the equipment, the length of the task queue and the geographical distribution of the collection points;

[0038] The priority scoring function is as follows:

[0039]

[0040] Among them, Ci is the task urgency, Ei is the energy consumption estimate, and Di is the geographical distance.

[0041] Preferably, the three-dimensional model supports interaction with a desktop, mobile terminal or augmented reality platform via WebGL, including browsing, annotation and measurement functions.

[0042] Preferably, the decoder can be connected to the Internet of Things sensor, and real-time data such as temperature, humidity and noise are mapped to the model node through spatial binding to achieve visual display of spatial data.

[0043] Preferably, the method is deployed in a distributed microservice architecture based on container orchestration, supports multi-threaded modeling, task fault tolerance and remote scheduling, and is suitable for large-scale urban modeling fields.

[0044] Beneficial effects

[0045] The present invention provides a method for constructing a real-life three-dimensional scene. Compared with the prior art, it has the following advantages:

[0046] Beneficial effects:

[0047] 1. In this invention, by introducing an improved iterative closest point algorithm, high-precision rigid registration between lidar point cloud data and image and geographic information data is achieved. On the basis of the traditional ICP algorithm, a dynamic weight adjustment mechanism is adopted, and a weighted matrix is introduced into the objective function of minimizing the registration residual. The registration contribution of different data sources can be adaptively adjusted according to their quality, density and spatial distribution characteristics, effectively improving the spatial fusion accuracy between heterogeneous data and reducing the risk of model offset or distortion caused by sensor errors. On this basis, the multidimensional data set in a unified coordinate system formed by the fusion is input into a semantic segmentation model based on a convolutional neural network. The model adopts The adaptive attention mechanism extracts and enhances the spatial features of the image. The encoder captures information at different spatial levels through a multi-scale convolution module, while the decoder introduces channel attention and spatial attention modules to effectively improve the segmentation accuracy of small targets, complex contours, and texture fuzzy areas. Cross-layer semantic fusion is achieved through residual connections, further improving the model's performance in multi-class target recognition and boundary preservation. Compared with existing modeling technologies, the present invention not only achieves accurate reconstruction of structural geometry, but also has strong semantic expression capabilities, which is beneficial to subsequent automatic labeling, scene recognition, and intelligent analysis tasks, significantly improving the intelligence and practicality of the three-dimensional modeling system.

[0048] 2. In the present invention, the real-scene 3D construction method introduces dynamic texture mapping technology in the model rendering stage, and combines the perspective heat map and historical trajectory prediction mechanism to realize the intelligent scheduling and loading optimization of texture data. The high-frequency detail area adopts a multi-level detail control mechanism, which can dynamically adjust the resolution of the loaded texture according to the user's current observation position and perspective distance, thereby reducing the graphics processing burden and system delay while ensuring visual quality. For low-frequency areas, this method introduces a sliding window prediction mechanism, records and analyzes the user's perspective direction vector sequence of the past N frames, uses a linear prediction model to estimate the perspective of the next frame, and then combines the hot spot area distribution map to preload the texture data in the predicted area. It significantly reduces texture loading delays and flickering when switching perspectives. At the same time, the modeling process adopts an incremental modeling mechanism driven by time series changes, which can perform time series comparison and difference analysis on data collected in consecutive time periods, identify newly appearing, moving or disappearing objects in the scene, and quickly update the model through hash matching and sparse difference modeling modules. Combined with the multi-platform deployment capability of WebGL, the constructed 3D model can achieve seamless interactive operations on desktop, mobile and AR platforms, including browsing, annotation, measurement, editing and other functions, fully meeting the application needs of urban management, emergency response, smart transportation and other applications, and significantly enhancing the real-time performance, scalability and user interaction experience of the model system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flow chart of a method for constructing a real-life three-dimensional scene proposed by the present invention;

[0050] Figure 2 A schematic diagram of experimental data for a method of constructing a real-life three-dimensional scene proposed in the present invention;

[0051] Figure 3 This is a time comparison view of multiple scene modeling of a method for constructing a real-life 3D scene proposed by the present invention;

[0052] Figure 4 This is a schematic diagram of the change in rendering frame rate over time in a method for constructing a real-life 3D scene proposed by the present invention;

[0053] Figure 5 A schematic diagram of the comprehensive performance radar of the method for constructing a real-life three-dimensional scene proposed by the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] See also Figure 1-Figure 5 , the present invention provides a technical solution, specifically including the following embodiments:

[0056] Example:

[0057] A method for constructing a real-life three-dimensional scene, comprising the following steps:

[0058] S1. Collect multi-source heterogeneous data of the scene to be constructed, including image data, lidar point cloud data and geographic information data;

[0059] S2, spatially register and fuse multi-source data to form a multidimensional data set in a unified coordinate system;

[0060] S3: Build an initial 3D model based on the fused dataset and perform semantic segmentation and classification on the model;

[0061] S4, using dynamic texture mapping technology to perform multi-scale real-time rendering of 3D models;

[0062] S5. The constructed three-dimensional scene supports multi-terminal access, real-time interaction and intelligent analysis applications.

[0063] The spatial registration in step S2 includes:

[0064] The point cloud data is rigidly registered based on the improved iterative closest point algorithm, in which a weighting mechanism is introduced to optimize the fusion accuracy between different data sources;

[0065] The registration process is achieved by minimizing the weighted residual sum of squares objective function, which is formulated as follows:

[0066]

[0067] Among them, wi is the dynamic weight, pi and qi are the corresponding point pairs in the source point cloud and the target point cloud respectively.

[0068] R is the rotation matrix and t is the translation vector.

[0069] It should be noted that the detailed process of spatial registration is:

[0070] Data Alignment and Initialization: In the initial registration phase, we first determine the source and target point cloud datasets. The source point cloud is the data to be registered, while the target point cloud serves as the reference standard. During the registration process, we initialize the position and pose of the source point cloud (i.e., rotation matrix and translation vector) and perform a preliminary coarse registration using the corresponding geometric transformation model.

[0071] Improved Iterative Closest Point Algorithm: The traditional ICP algorithm finds the closest point pair between the source and target point clouds and minimizes the error based on this. To improve registration accuracy, this invention employs a weighting mechanism to optimize the registration accuracy between different data sources. This weighting mechanism helps address issues such as scale differences and noise interference that may exist when registering different data sources, resulting in a more accurate registration result.

[0072] During this process, through iterative optimization, each point in the source point cloud continuously searches for the closest point in the target point cloud and calculates the rotation matrix and translation vector to achieve alignment. Each iteration gradually reduces the registration error until the error converges to a certain threshold.

[0073] Semantic segmentation in S3 uses a convolutional neural network with an adaptive attention mechanism, which includes:

[0074] The encoder uses multi-scale convolution to extract spatial features, and the decoder introduces channel and spatial attention modules to improve boundary recognition.

[0075] Cross-layer feature fusion is achieved through residual connections to enhance the segmentation accuracy of small targets and complex backgrounds.

[0076] It should be noted that the role of the encoder part is to extract spatial features from the input data. In order to improve the extraction capability, multi-scale convolution is used. Multi-scale convolution uses convolution kernels of different sizes to capture spatial information of different scales, ensuring that the network can understand various features in the scene from coarse to fine. By using convolution kernels of different sizes, the network can effectively extract local features of different sizes. For example, larger convolution kernels help capture large-scale structures, while smaller convolution kernels focus on capturing detailed information, which is crucial for complex three-dimensional scenes. The encoder gradually compresses spatial information and extracts feature maps through multi-layer convolution operations. These feature maps include The decoder contains the semantic information of the input data for the subsequent decoding process. The decoder is responsible for restoring the features extracted by the encoder into the final segmentation result. In order to solve the problem of unclear boundaries in complex scenes, the channel and spatial attention modules are introduced in the decoder to enhance the focus on key areas (such as boundary areas) and details. When performing semantic segmentation, especially when facing small targets and complex backgrounds, detailed information is often lost, affecting the segmentation accuracy. To alleviate this problem, we use residual connections in the network. Residual connections allow information to be directly transmitted between different layers, thereby maintaining the details of low-level features while avoiding gradient disappearance and information loss.

[0077] Dynamic texture mapping in S4 includes:

[0078] High-frequency areas use a level-of-detail control mechanism to adjust texture resolution in real time based on viewing distance;

[0079] Texture caching and heat prediction mechanisms are introduced in low-frequency areas to improve rendering efficiency.

[0080] It should be noted that:

[0081] The texture prediction mechanism uses a sliding window algorithm based on the user's viewing history to predict the viewing angle change trend and preload the texture of the target area. Specifically, it includes:

[0082] Record the user perspective vector sequence {ut-N+1,.....,ut} for the past N frames;

[0083] Use the linear prediction model to estimate the viewing direction of the next frame. The formula is as follows:

[0084]

[0085] Among them, α∈(0,1).

[0086] It should be noted that high-frequency areas usually refer to the details in the model, such as the edges of objects or important structures. In these areas, higher texture resolution is required to ensure clear details. To achieve this goal, the system adjusts the texture resolution in real time according to the observation distance. The specific approach is as follows: when the user approaches the high-frequency area, the system will automatically increase the texture resolution of the area to ensure the clarity of the details. Conversely, when the user is away from the area, the system will reduce the texture resolution to reduce the computational burden. This mechanism is similar to the level of detail (LOD) technology. Through dynamic adjustment of different levels of detail (such as multiple texture layers from coarse to fine), the system can select the appropriate texture resolution according to demand. Low-frequency areas usually refer to areas in the scene that are less conspicuous or far away from the observer. The texture resolution requirements of these areas are relatively low. In order to improve rendering efficiency, the system introduces texture caching and heat prediction mechanisms in low-frequency areas. The specific mechanisms are as follows:

[0087] Texture cache: To avoid repeated loading of textures, the system stores the already loaded textures in the cache. When the user's perspective approaches a low-frequency area, the texture is directly obtained from the cache, avoiding repeated loading and reducing rendering delays.

[0088] Heat prediction mechanism: By analyzing historical data of user perspectives, the system can predict which areas will soon enter the field of view and load the textures of the corresponding areas in advance. For example, if the user's perspective history shows that they are about to view a certain area, the system will preload the texture data of that area based on this trend, reducing waiting time.

[0089] During the construction process, an incremental modeling mechanism based on time series changes is introduced, including:

[0090] Compare the data collected at different time periods in time to detect the addition, movement and deletion of objects;

[0091] Use the differential modeling module to update the 3D model and update the sparse difference extraction strategy based on hash matching.

[0092] It should be noted that during the construction of a three-dimensional scene, the collected data will change over time. In order to achieve dynamic updates, the system will perform a time series comparison on the data collected at different times to detect which objects have changed, mainly including the following types of changes.

[0093] New objects: When new objects or structures appear in the collected data over a certain period of time, the system will identify these new objects and add them to the 3D model.

[0094] Moving objects: When the position of an existing object changes, the system detects the object's displacement and updates its position in the model.

[0095] Deleted objects: When certain objects disappear from the data for a certain period of time, the system will mark these objects as deleted and remove them from the 3D model.

[0096] Update mechanism of differential modeling module: The differential modeling module extracts the differences in model changes by comparing 3D data from different time periods, and only updates these changes. This approach can significantly reduce the amount of data that needs to be processed during each update and improve modeling efficiency.

[0097] Multi-source heterogeneous data collection uses an asynchronous collection scheduling mechanism, including:

[0098] Adopt energy consumption priority scheduling strategy based on the remaining power of the equipment, the length of the task queue and the geographical distribution of the collection points;

[0099] The priority scoring function is as follows:

[0100]

[0101] Among them, Ci is the task urgency, Ei is the energy consumption estimate, and Di is the geographical distance.

[0102] It should be noted that:

[0103] The 3D model supports interaction with desktop, mobile terminals or augmented reality platforms through WebGL, including browsing, annotation and measurement functions.

[0104] It should be noted that when collecting data from multiple sources, the energy consumption of different devices is an important consideration. The priority of each device's collection task is adjusted based on its remaining power, thereby ensuring efficient energy use of the device and avoiding interruptions to collection tasks due to insufficient power. The scheduling strategy includes the following aspects:

[0105] Remaining power evaluation: The remaining power of a device will affect the scheduling priority of its tasks. If the device's power is low, the system will prioritize assigning tasks to devices with sufficient power.

[0106] Task queue length: The length of the task queue determines the number of tasks currently to be processed. The system will adjust the collection order of tasks based on the queue length to ensure that urgent tasks are processed first.

[0107] Geographic distribution: The geographical distribution of collection points will affect the execution order of tasks. For collection points with a wide distribution, the system will optimize task scheduling based on distance to reduce the movement time and energy consumption of equipment.

[0108] The decoder can be connected to IoT sensors, and real-time data such as temperature, humidity and noise are mapped to model nodes through spatial binding to achieve visual display of spatial data.

[0109] It is important to note that the decoder integrates with IoT sensors to obtain real-time environmental data. For example, a temperature sensor monitors the temperature of a specific area, a humidity sensor collects air humidity data, and a noise sensor detects the ambient noise level. This data is synchronized with the 3D modeling system in real time to provide accurate spatial information.

[0110] Spatial binding and data mapping: The acquired environmental data needs to be bound to the corresponding nodes in the 3D model. Through the spatial binding mechanism, the system can map the real-time data to a specific location or area of the model. This process includes:

[0111] Node Mapping: Each sensor data point is associated with a specific location node in the 3D model. For example, a temperature sensor may be tied to a location in a building, and a humidity sensor may be tied to a green area in the model.

[0112] Spatial association: Through the data provided by sensors, the system can determine the spatial relationship between various physical elements (such as walls, ground, vegetation, etc.) and environmental data, and then dynamically update the three-dimensional model.

[0113] This method is deployed in a distributed microservice architecture based on container orchestration, supports multi-threaded modeling, task fault tolerance and remote scheduling, and is suitable for large-scale urban modeling scenarios.

[0114] It should be noted that each module (such as data collection, data processing, 3D modeling, rendering, semantic segmentation, etc.) is deployed in the form of containers. Containerization can provide a lightweight and portable environment, so that each microservice can run, expand and manage independently. Container orchestration tools are used to manage and schedule containers. Container orchestration can achieve automated deployment, expansion, monitoring and load balancing to cope with the large amount of data processing and requests that may arise in urban modeling. The functions are split into multiple small, independent service modules. Each service module performs different functions, such as data collection, data processing, modeling, rendering, etc. Each microservice only focuses on a single task, which can better share the system load while improving the maintainability and scalability of the system.

[0115] Working principle: By fusing multi-source heterogeneous perception data and combining it with intelligent algorithms, high-precision reconstruction and semantic enhancement of three-dimensional space are achieved. First, the system collects a variety of heterogeneous data including images, lidar point clouds and geographic information, and performs spatial registration in a unified coordinate system. The improved iterative closest point (ICP) algorithm is used in the registration process, and a weighted residual minimization mechanism is introduced to adaptively adjust the fusion weights of different data sources, thereby improving the rigid alignment accuracy between point clouds and images. After the fusion is completed, the system calls the convolutional neural network model with an adaptive attention mechanism for semantic segmentation and classification based on the constructed multidimensional data set. The network extracts spatial features through multi-scale convolution, and introduces channel and spatial attention modules in the decoding stage to enhance the recognition ability of target edges and small-sized targets. At the same time, cross-layer feature fusion is achieved with the help of residual connections to improve segmentation accuracy and semantic consistency. Subsequently, after the three-dimensional modeling is completed, the system performs dynamic segmentation on the model. Dynamic texture mapping operation adopts a partitioning strategy to take differentiated treatments on areas of different visual importance: the high-frequency area adopts the level of detail control (LOD) mechanism based on the user's perspective and distance to adjust the texture resolution in real time, and the low-frequency area introduces a sliding window prediction algorithm. By analyzing the historical sequence of user perspective and combining it with the heat map, the future observation area is predicted, and the corresponding texture resources are loaded in advance to reduce rendering latency. To achieve continuous scene updates, the present invention further introduces an incremental modeling mechanism based on time series changes. The addition, movement and deletion of scene objects are detected through time series data frame comparison and sparse difference recognition algorithm, and the local reconstruction and update of the three-dimensional model are quickly completed. The final constructed model supports cross-platform access through WebGL, including desktop, mobile and augmented reality devices, and supports linkage with IoT devices to visualize environmental sensor data in real time. It is widely applicable to various complex scenarios such as urban modeling, emergency response, and intelligent security.

[0116] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the application should be included in the scope of protection of the present application.

Claims

1. A method for constructing a real-life three-dimensional scene, characterized by: The following steps are involved: S1. Collect multi-source heterogeneous data of the scene to be constructed, including image data, lidar point cloud data and geographic information data; S2, spatially register and fuse multi-source data to form a multidimensional data set in a unified coordinate system; S3: Build an initial 3D model based on the fused dataset and perform semantic segmentation and classification on the model; S4, using dynamic texture mapping technology to perform multi-scale real-time rendering of 3D models; S5. The constructed three-dimensional scene supports multi-terminal access, real-time interaction and intelligent analysis applications.

2. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: The spatial registration in step S2 includes: The point cloud data is rigidly registered based on the improved iterative closest point algorithm, in which a weighting mechanism is introduced to optimize the fusion accuracy between different data sources; The registration process is achieved by minimizing the weighted residual square sum objective function, which is formulated as follows: Among them, wi is the dynamic weight, pi and qi are the corresponding point pairs in the source point cloud and the target point cloud respectively. R is the rotation matrix and t is the translation vector.

3. The method for constructing a real-life three-dimensional scene according to claim 2, wherein: The semantic segmentation in S3 uses a convolutional neural network with an adaptive attention mechanism, specifically including: The encoder uses multi-scale convolution to extract spatial features, and the decoder introduces channel and spatial attention modules to improve boundary recognition. Cross-layer feature fusion is achieved through residual connections to enhance the segmentation accuracy of small targets and complex backgrounds.

4. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: The dynamic texture mapping in S4 includes: High-frequency areas use a level-of-detail control mechanism to adjust texture resolution in real time based on viewing distance; Texture caching and heat prediction mechanisms are introduced in low-frequency areas to improve rendering efficiency.

5. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: The texture prediction mechanism uses a sliding window algorithm based on the user's viewing history to predict the viewing angle change trend and preload the target area texture. Specifically, it includes: Record the user perspective vector sequence {ut-N+1,.....,ut} for the past N frames; Use the linear prediction model to estimate the viewing direction of the next frame. The formula is as follows: Among them, α∈(0,1).

6. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: During the construction process, an incremental modeling mechanism based on time series changes is introduced, including: Compare the data collected at different time periods in time to detect the addition, movement and deletion of objects; The 3D model is updated using a differential modeling module, and the update is based on a sparse difference extraction strategy using hash matching.

7. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: The multi-source heterogeneous data collection adopts an asynchronous collection scheduling mechanism, including: Adopt energy consumption priority scheduling strategy based on the remaining power of the equipment, the length of the task queue and the geographical distribution of the collection points; The priority scoring function is as follows: Among them, Ci is the task urgency, Ei is the energy consumption estimate, and Di is the geographical distance.

8. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: The three-dimensional model supports interaction with desktop, mobile terminals or augmented reality platforms through WebGL, including browsing, annotation and measurement functions.

9. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: The decoder can be connected to IoT sensors, and real-time data such as temperature, humidity and noise are mapped to model nodes through spatial binding to achieve visual display of spatial data.

10. The method for constructing a real-life three-dimensional scene according to claim 1, wherein: This method is deployed in a distributed microservice architecture based on container orchestration, supports multi-threaded modeling, task fault tolerance and remote scheduling, and is suitable for large-scale urban modeling scenarios.

Images