An unreal engine-based MR virtual scene intelligent construction method and system

Abstract

The application provides a kind of based on unreal engine MR virtual scene intelligent construction method and system, it is related to MR virtual scene construction technical field, including: multi-sensor fusion module, artificial intelligence analysis module, unreal engine scene construction module, MR interaction module and continuous optimization module, each module works cooperatively, realize the intelligent, automation construction of MR virtual scene;The application breaks through the limitation that traditional MR scene construction relies on artificial modeling and manual configuration, automatically collects multi-source environmental data through the multi-sensor fusion module, combines the preprocessing, semantic segmentation and intelligent generation capability of the artificial intelligence analysis module, realizes the full-process automation from data acquisition to virtual content generation, significantly reduces the cost of artificial intervention, solves the problems of traditional MR virtual scene construction mode, such as relying on manual operation, long construction period, insufficient semantic analysis and virtual-real fusion precision, low scene adaptation flexibility and lack of continuous optimization mechanism.

Description

Technical Field

[0001] This invention relates to the field of MR virtual scene construction technology, and in particular to an intelligent method and system for constructing MR virtual scenes based on Unreal Engine. Background Technology

[0002] With the rapid development of mixed reality (MR) technology, it has been widely used in various fields such as surveying and mapping, architectural engineering, urban planning, and natural landform exploration due to its characteristic of blending virtual and real elements. As the core carrier, MR virtual scenes need to achieve accurate matching, high-fidelity presentation, and natural interaction between virtual content and the real environment in order to meet the needs of practical applications for scene immersion, data accuracy, and ease of operation.

[0003] However, traditional MR virtual scene construction methods have many limitations, which seriously restrict the large-scale implementation and efficiency improvement of the technology: Relying on manual operation and incurring high construction costs: Traditional scene construction requires professional technicians to manually create virtual object models, design scene layouts, and manually configure rules such as lighting and physical collision, as well as MR interaction interfaces, using 3D modeling software. The entire process demands extremely high professional skills from personnel and is cumbersome and time-consuming. Especially when facing complex surveying scenarios such as urban terrain and large-scale construction projects, the workload of manual modeling and configuration increases exponentially, resulting in persistently high scene construction costs.

[0004] The construction cycle is lengthy and the response efficiency is low: Due to excessive human intervention, the entire process, from data collection and model building to scene debugging and optimization, often takes weeks or even months to complete. For scenarios that require rapid iteration or urgent response (such as emergency mapping, temporary engineering planning, etc.), traditional methods are unable to meet the timeliness requirements and cannot provide usable MR virtual scenes in a timely manner.

[0005] Insufficient semantic parsing accuracy and poor virtual-real fusion: In traditional scene construction, semantic understanding of the real environment relies heavily on manual annotation, which is not only inefficient but also prone to annotation errors. Furthermore, the lack of specialized semantic parsing tools for the surveying and mapping geographic information field makes it difficult to accurately identify key features such as terrain and land features, resulting in low matching between the generated virtual content and the real environment. In addition, spatial alignment between the virtual scene and the real environment depends on manual calibration, leading to significant pose deviations, typically at the centimeter level or even higher, severely impacting the user's immersive experience and the accuracy of data application.

[0006] Poor scene adaptability and insufficient flexibility: Traditional MR scene building systems are mostly customized, with fixed scene parameters, virtual content types, and interaction rules, making it difficult to adapt to the differentiated needs of different surveying and mapping scenarios (such as urban terrain, building engineering, natural landforms, etc.). If it is necessary to switch application scenarios, large-scale system reconstruction and manual adjustments are often required, which is time-consuming, labor-intensive, and has poor compatibility.

[0007] The lack of a continuous optimization mechanism hinders user experience improvement: Once traditional scenarios are built, their effects and interaction logic are essentially fixed, allowing only limited adjustments through manual maintenance. Because real-time user interaction data cannot be collected and intelligently analyzed, it's difficult to accurately identify problems within the scenario (such as unreasonable virtual object placement, delayed interaction responses, and inconsistent visual effects). This prevents the scenario experience from being continuously iterated and optimized based on actual user feedback, making it difficult to meet long-term user needs.

[0008] Therefore, developing a technical solution that can realize intelligent and automated construction of MR virtual scenes, reduce labor costs, shorten the construction cycle, improve scene adaptability and virtual-real fusion accuracy, and has continuous optimization capabilities has become a key issue that urgently needs to be addressed in the current field of MR virtual scene construction. Summary of the Invention

[0009] This invention relates to an intelligent construction method and system for MR virtual scenes based on Unreal Engine, which solves the problems of traditional MR virtual scene construction methods, such as reliance on manual operation, long construction cycle, insufficient accuracy of semantic parsing and virtual-real fusion, low flexibility of scene adaptation, and lack of continuous optimization mechanism.

[0010] This invention provides an intelligent MR virtual scene construction system based on Unreal Engine, specifically including: a multi-sensor fusion module, an artificial intelligence analysis module, an Unreal Engine scene construction module, an MR interaction module, and a continuous optimization module. These modules work collaboratively to achieve intelligent and automated construction of MR virtual scenes. The multi-sensor fusion module is connected to a LiDAR, a high-definition camera, and an inertial measurement unit, respectively, to integrate multi-source data collected by these three types of sensors and corresponding environmental semantic information, providing high-quality data support for subsequent scene construction. The artificial intelligence analysis module includes a data preprocessing submodule, a semantic segmentation submodule, and a scene generation submodule, and is communicatively connected to the multi-sensor fusion module, used for preprocessing multi-source data... The system includes semantic parsing and intelligent generation of virtual scene-related content. The Unreal Engine scene building module is electrically connected to the AI ​​analysis module, receiving the generated virtual content data, building a basic scene based on real-world parameters, and performing high-fidelity visualization rendering of the virtual content. It also configures relevant rules and interfaces for MR interaction. The MR interaction module includes MR glasses and an interactive controller, interacting with the Unreal Engine scene building module to display the virtual scene, receive user operations, and collect real-time pose tracking data. The continuous optimization module is electrically connected to the MR interaction module, the Unreal Engine scene building module, and the AI ​​analysis module, respectively, to provide user interaction data feedback and transmit optimization commands, iterating through algorithms to optimize scene effects and interactive experience.

[0011] Furthermore, the multi-sensor fusion module acquires 3D point cloud data of the real environment through LiDAR, RGB image data and depth image data through a high-definition camera, and device pose data through an inertial measurement unit. At the same time, it integrates the environmental semantic information corresponding to the above data, including object category, spatial location and topological relationship.

[0012] Furthermore, the semantic segmentation submodule of the artificial intelligence analysis module adopts an improved MaskR-CNN model, which adds a branch for extracting geographic information semantic features. This model can adapt to scene semantic parsing in the field of surveying and mapping geographic information and accurately identify features such as terrain and land features.

[0013] Furthermore, the scene generation submodule of the artificial intelligence analysis module adopts a pre-trained generative adversarial network model, which includes a generator and a discriminator. The generator is used to generate virtual object models, scene layout schemes and interaction logic, and the discriminator is used to determine the matching degree between the generated content and the real environment. The generated content is optimized through adversarial training.

[0014] Furthermore, the Unreal Engine scene building module is used to create a basic scene framework of equal scale based on the scale parameters of the real environment, configure the scene's lighting system, physical collision rules and interactive interfaces adapted to the MR interaction module, and at the same time perform high-fidelity visualization rendering of the virtual content generated by the scene generation submodule through the real-time rendering module.

[0015] Furthermore, the MR glasses of the MR interaction module are used to display virtual scenes and collect real-time pose tracking data, and the interactive handle is used to receive user operation commands and transmit them to the Unreal Engine scene building module to realize real-time interaction between the user and the MR scene.

[0016] Furthermore, the continuous optimization module collects user interaction data fed back by the MR interaction module, uses reinforcement learning algorithms to generate optimization instructions, and transmits them to the artificial intelligence analysis module and the Unreal Engine scene construction module respectively, to iteratively optimize the position, shape, interaction response mechanism and scene visual effects of virtual objects.

[0017] Furthermore, the data preprocessing submodule of the artificial intelligence analysis module is used to preprocess the multi-source data transmitted by the multi-sensor fusion module, including point cloud denoising, registration, downsampling, as well as image deblurring, enhancement, and feature extraction operations.

[0018] Furthermore, the Unreal Engine scene building module and MR interaction module achieve precise spatial alignment between the virtual scene and the real environment through a coordinate transformation algorithm. The coordinate transformation algorithm uses Kalman filtering for pose optimization to improve alignment accuracy.

[0019] A method for constructing an intelligent MR virtual scene construction system based on Unreal Engine includes the following steps: Step 1. Activate the multi-sensor fusion module to collect 3D point cloud data, RGB image data, depth image data, device pose data, and environmental semantic information of the real environment through LiDAR, high-definition camera, and inertial measurement unit; Step 2. The multi-sensor fusion module transmits the collected multi-source data to the artificial intelligence analysis module, and the data preprocessing submodule completes the data preprocessing operation. Step 3. The preprocessed data undergoes semantic parsing by the semantic segmentation submodule to generate a structured environment semantic model containing object semantic labels; Step 4. The environmental semantic model input scene generation submodule automatically generates virtual object models, scene layout schemes, and interaction logic through a generative adversarial network model; Step 5. The Unreal Engine scene building module receives the above virtual content, builds a basic scene framework in proportion to the real environment, and configures the lighting system, physical collision rules and interaction interfaces; Step 6. The Unreal Engine scene building module achieves high-fidelity visualization of the virtual scene through the real-time rendering module and establishes communication with the MR interaction module; Step 7. The user views the virtual scene through MR glasses and interacts with the interactive controller. The MR interaction module collects real-time pose tracking data and user operation feedback. Step 8. The continuous optimization module receives data transmitted by the MR interaction module and uses reinforcement learning algorithms to generate optimization instructions; Step 9. Adjust the generation strategy of the artificial intelligence analysis module and the scene parameters of the Unreal Engine scene building module according to the optimization instructions to complete the iterative optimization.

[0020] This invention provides an intelligent method and system for constructing MR virtual scenes based on Unreal Engine, which has the following beneficial effects: This invention breaks through the limitations of traditional MR scene construction, which relies on manual modeling and configuration. By automatically collecting multi-source environmental data through a multi-sensor fusion module and combining the preprocessing, semantic segmentation, and intelligent generation capabilities of an artificial intelligence analysis module, it achieves full automation from data collection to virtual content generation, significantly reducing the cost of manual intervention and shortening the scene construction cycle by more than 50%.

[0021] The semantic segmentation submodule of this invention adopts an improved MaskR-CNN model and adds a branch for extracting geographic information semantic features. It achieves accurate identification of features such as terrain and land features in the field of surveying and mapping geographic information, and the semantic parsing accuracy is improved to over 90%, avoiding errors and tedious operations caused by manual annotation. The scene generation submodule automatically generates virtual object models, layout schemes and interaction logic that are adapted to the real environment through a pre-trained generative adversarial network model. The generated content has a high degree of matching with the real environment and does not require repeated manual adjustments.

[0022] The multi-sensor fusion module of this invention integrates 3D point cloud data from LiDAR, RGB and depth image data from high-definition cameras, and device pose data from inertial measurement units. It also associates semantic information such as object category, spatial location, and topological relationship, providing multi-dimensional and high-quality data support for scene construction and laying the foundation for accurate virtual-real fusion.

[0023] This invention's Unreal Engine scene building module creates a proportionally scaled basic scene based on real-world environmental scale parameters, configures a lighting system and collision rules that conform to real-world physical laws, and combines high-fidelity real-time rendering technology to ensure that the visual effects and physical properties of virtual content are highly consistent with the real environment. Through a coordinate transformation algorithm optimized by Kalman filtering, it achieves precise spatial alignment between the virtual scene and the real environment, with pose deviation controlled at the millimeter level, significantly improving the user's immersive experience.

[0024] The MR interaction module of this invention achieves high-definition display of virtual scenes, real-time reception of user operations, and accurate acquisition of pose data through the collaboration of MR glasses and interactive handles. The interaction response latency is less than 10ms, and the operation logic conforms to user habits, ensuring the smoothness and naturalness of human-computer interaction.

[0025] The system of this invention supports adaptation to different surveying and mapping scenarios (such as urban terrain, building engineering, natural landforms, etc.). Through the model generalization capability of the artificial intelligence analysis module and the scene configuration flexibility of Unreal Engine, scene parameters, virtual content types and interaction rules can be quickly adjusted to meet diverse application needs without the need for large-scale system reconstruction.

[0026] The continuous optimization module of this invention collects user interaction data and uses reinforcement learning algorithms to generate targeted optimization instructions, which are then fed back to the artificial intelligence analysis module and the Unreal Engine scene construction module. This enables iterative optimization of the virtual object's position, shape, interaction response mechanism, and scene visual effects, allowing the system to continuously evolve based on actual user feedback. Moreover, the optimization process requires no manual intervention; the algorithm automatically completes parameter adjustments and strategy updates, ensuring both the timeliness and accuracy of optimization while reducing later maintenance costs. This allows the MR virtual scene to maintain the optimal user experience in the long term. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0028] The accompanying drawings described below are only related to some embodiments of the invention and are not intended to limit the invention.

[0029] In the attached diagram: Figure 1 The overall system architecture block diagram of the present invention is shown; Figure 2 A schematic diagram of the improved Mask R-CNN model structure of the present invention is shown; Figure 3 This diagram illustrates the adversarial training process of the GAN model of the present invention. Figure 4 A schematic diagram of the method flow of the present invention is shown; List of reference numerals 1. Multi-sensor fusion module; 101. LiDAR; 102. High-definition camera; 103. Inertial measurement unit; 2. Artificial intelligence analysis module; 201. Data preprocessing submodule; 202. Semantic segmentation submodule; 203. Scene generation submodule; 3. Unreal Engine scene construction module; 4. MR interaction module; 401. MR glasses; 402. Interactive handle; 5. Continuous optimization module. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the described embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example: Please refer to Figures 1 to 4 : This invention proposes an intelligent MR virtual scene construction system based on Unreal Engine, comprising: a multi-sensor fusion module 1, an artificial intelligence analysis module 2, an Unreal Engine scene construction module 3, an MR interaction module 4, and a continuous optimization module 5. These modules work collaboratively to achieve intelligent and automated construction of MR virtual scenes. The multi-sensor fusion module 1 is connected to a LiDAR 101, a high-definition camera 102, and an inertial measurement unit 103, respectively, to integrate multi-source data collected by the three types of sensors and corresponding environmental semantic information, providing high-quality data support for subsequent scene construction. The multi-sensor fusion module 1 collects 3D point cloud data of the real environment through the LiDAR 101 and collects MR interaction data through the high-definition camera 102. GB image data and depth image data are collected by the inertial measurement unit 103 to acquire device pose data. Simultaneously, the environmental semantic information corresponding to the above data is integrated, including object category, spatial location, and topological relationships. The artificial intelligence analysis module 2 includes a data preprocessing submodule 201, a semantic segmentation submodule 202, and a scene generation submodule 203, and is communicatively connected to the multi-sensor fusion module 1. It is used for preprocessing multi-source data, semantic parsing, and intelligent generation of virtual scene-related content. The semantic segmentation submodule 202 of the artificial intelligence analysis module 2 adopts an improved Mask R-CNN model, which adds a branch for extracting geographic information semantic features, enabling it to adapt to the field of surveying and mapping geographic information. Scene semantic parsing accurately identifies features such as terrain and land features; the scene generation submodule 203 of the AI ​​analysis module 2 adopts a pre-trained generative adversarial network model, which includes a generator and a discriminator. The generator is used to generate virtual object models, scene layout schemes, and interaction logic, while the discriminator is used to determine the matching degree between the generated content and the real environment. The generated content is optimized through adversarial training; the data preprocessing submodule 201 of the AI ​​analysis module 2 is used to preprocess the multi-source data transmitted by the multi-sensor fusion module 1, including point cloud denoising, registration, downsampling, as well as image deblurring, enhancement, and feature extraction operations; the Unreal Engine scene construction module 3 is electrically connected to the AI ​​analysis module 2 and receives its... The generated virtual content data is used to build a basic scene based on real environment parameters and complete high-fidelity visualization rendering of the virtual content. At the same time, relevant rules and interfaces adapted to MR interaction are configured. The Unreal Engine scene building module 3 is used to create a proportional basic scene framework based on the scale parameters of the real environment, configure the scene's lighting system, physical collision rules and interactive interfaces adapted to the MR interaction module 4, and perform high-fidelity visualization rendering of the virtual content generated by the scene generation submodule 203 through the real-time rendering module. The MR interaction module 4 includes MR glasses 401 and interactive handles 402, which interact with the Unreal Engine scene building module 3 to realize virtual scene display, user operation reception and real-time pose tracking data acquisition.The MR glasses 401 of the MR interaction module 4 are used to display virtual scenes and collect real-time pose tracking data. The interactive handle 402 is used to receive user operation commands and transmit them to the Unreal Engine scene construction module 3, enabling real-time interaction between the user and the MR scene. The Unreal Engine scene construction module 3 and the MR interaction module 4 achieve precise spatial alignment between the virtual scene and the real environment through a coordinate transformation algorithm. The coordinate transformation algorithm uses Kalman filtering for pose optimization to improve alignment accuracy. The continuous optimization module 5 is electrically connected to the MR interaction module 4, the Unreal Engine scene construction module 3, and the artificial intelligence analysis module 2, respectively, to realize user interaction data feedback and optimization command transmission. Through algorithm iteration, it optimizes scene effects and interactive experience. The continuous optimization module 5 collects user interaction data fed back by the MR interaction module 4, uses reinforcement learning algorithms to generate optimization commands, and transmits them to the artificial intelligence analysis module 2 and the Unreal Engine scene construction module 3, respectively, to iteratively optimize the position, shape, interactive response mechanism, and scene visual effects of virtual objects.

[0032] The working principle of this embodiment: The multi-sensor fusion module 1 serves as the core data input, working collaboratively with three types of devices—LiDAR 101, HD camera 102, and inertial measurement unit 103—to comprehensively collect real-world environmental information. LiDAR 101 captures 3D point cloud data, accurately reconstructing the spatial geometry of the environment. HD camera 102 simultaneously acquires RGB and depth image data, supplementing the environment's color texture and object spacing information. Inertial measurement unit 103 records its own pose data in real time, providing a reference for subsequent spatial alignment. The multi-sensor fusion module 1 integrates the three types of raw data and associates them with environmental semantic information such as object category, spatial location, and topological relationships. It filters data redundancy, corrects acquisition deviations, and forms high-quality, structured multi-source fusion data, providing reliable support for subsequent intelligent analysis and scene construction. The AI ​​analysis module 2 receives multi-source fused data and completes a progressive processing of "data processing - semantic cognition - content generation" through three sub-modules. First, the data preprocessing sub-module 201 performs denoising, registration, and downsampling operations on point cloud data, and deblurring, enhancement, and feature extraction on image data to eliminate interference factors during the acquisition process and improve data quality. Second, the semantic segmentation sub-module 202 adopts an improved MaskR-CNN model and accurately analyzes key features such as terrain and land features in the field of surveying and mapping geographic information through a newly added geographic information semantic feature extraction branch, generating a structured environmental semantic model with semantic labels to achieve intelligent cognition of the real environment. Finally, the scene generation sub-module 203 is based on a pre-trained generative adversarial network (GAN). Through adversarial training between the generator and the discriminator, it automatically generates virtual object models, scene layout schemes, and interaction logic adapted to the real environment. The discriminator continuously verifies the matching degree between the generated content and the real environment, prompting the generator to optimize the output results and ensure the rationality and adaptability of the virtual content. The Unreal Engine scene building module 3 receives virtual content data output from the AI ​​analysis module 2 and builds a 1:1 scale basic scene framework based on the scale parameters of the real environment to ensure that the spatial scale of the virtual scene is consistent with that of the real environment. The Unreal Engine scene building module 3 simultaneously configures the scene's lighting system to simulate realistic lighting effects; sets physical collision rules to ensure that the interaction between virtual and real objects conforms to physical laws; and builds an interface adapted for MR interaction, paving the way for subsequent user interaction. Simultaneously, the engine's real-time rendering module performs high-fidelity visualization rendering of virtual objects, scene layout, and other content, restoring delicate textures, materials, and dynamic effects to generate an immersive initial MR virtual scene. The MR interaction module 4 serves as the human-computer interaction hub, enabling bidirectional data transmission through the MR glasses 401 and the interactive handle 402. The MR glasses 401 are responsible for displaying high-fidelity virtual scenes while simultaneously collecting real-time user pose tracking data, capturing information such as user movement and perspective changes in real space. The interactive handle 402 receives user operation commands (such as clicking, dragging, zooming, etc.) and transmits the commands to the Unreal Engine scene construction module 3, enabling users to control the virtual scene in real time. The MR interaction module 4 synchronously transmits user pose data and operation feedback data to the continuous optimization module 5, providing a basis for real user needs to optimize the scene. As the core of the system's closed loop, the continuous optimization module 5 receives user pose data, operation behavior data, and experience feedback from the MR interaction module 4. It performs in-depth analysis of the data using reinforcement learning algorithms to identify issues such as virtual object position deviations, interaction response delays, and visual inconsistencies, generating targeted optimization instructions. These instructions are then transmitted to the artificial intelligence analysis module 2 and the Unreal Engine scene construction module 3: the former adjusts the virtual content generation strategy and optimizes the virtual object shapes and scene layout logic; the latter corrects scene lighting parameters, physical collision rules, rendering effects, and interaction interface configurations. Simultaneously, the Unreal Engine scene construction module 3 and the MR interaction module 4, using a coordinate transformation algorithm optimized by Kalman filtering and combined with user pose data collected by the MR glasses 401, achieve precise spatial alignment between the virtual scene and the real environment, reducing virtual-real discrepancies. Through multiple rounds of iterative "feedback-analysis-optimization" cycles, the accuracy of virtual-real fusion, visual immersion, and interactive smoothness of the MR virtual scene are continuously improved.

Claims

1. An intelligent MR virtual scene construction system based on Unreal Engine, characterized in that, include: The system comprises a multi-sensor fusion module (1), an artificial intelligence analysis module (2), an Unreal Engine scene construction module (3), an MR interaction module (4), and a continuous optimization module (5). These modules work together to achieve intelligent and automated construction of MR virtual scenes. The multi-sensor fusion module (1) is connected to a lidar (101), a high-definition camera (102), and an inertial measurement unit (103) to integrate multi-source data collected by the three types of sensors and corresponding environmental semantic information, providing high-quality data support for subsequent scene construction. The artificial intelligence analysis module (2) includes a data preprocessing submodule (201), a semantic segmentation submodule (202), and a scene generation submodule (203), and is communicatively connected to the multi-sensor fusion module (1) to perform preprocessing, semantic parsing, and virtualization of multi-source data. Intelligent generation of scene-related content; the Unreal Engine scene building module (3) is electrically connected to the artificial intelligence analysis module (2), receives the virtual content data generated by it, builds a basic scene based on real environment parameters and completes high-fidelity visualization rendering of virtual content, and configures relevant rules and interfaces for MR interaction; the MR interaction module (4) includes MR glasses (401) and an interactive handle (402), interacts with the Unreal Engine scene building module (3), and realizes virtual scene display, user operation reception and real-time pose tracking data collection; the continuous optimization module (5) is electrically connected to the MR interaction module (4), the Unreal Engine scene building module (3) and the artificial intelligence analysis module (2) respectively, realizes user interaction data feedback and optimization instruction transmission, and optimizes scene effects and interactive experience through algorithm iteration.

2. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 1, characterized in that, The multi-sensor fusion module (1) collects three-dimensional point cloud data of the real environment through lidar (101), collects RGB image data and depth image data through high-definition camera (102), collects device pose data through inertial measurement unit (103), and integrates the environmental semantic information corresponding to the above data. The semantic information includes object category, spatial location and topological relationship.

3. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 2, characterized in that, The semantic segmentation submodule (202) of the artificial intelligence analysis module (2) adopts the improved MaskR-CNN model, which adds a branch for extracting geographic information semantic features. It can adapt to the scene semantic analysis in the field of surveying and mapping geographic information and accurately identify features such as terrain and land features.

4. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 3, characterized in that, The scene generation submodule (203) of the artificial intelligence analysis module (2) adopts a pre-trained generative adversarial network model. The model includes a generator and a discriminator. The generator is used to generate virtual object models, scene layout schemes and interaction logic. The discriminator is used to judge the matching degree between the generated content and the real environment. The generated content is optimized through adversarial training.

5. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 4, characterized in that, The Unreal Engine scene building module (3) is used to create a basic scene framework of equal scale based on the scale parameters of the real environment, configure the scene's lighting system, physical collision rules and interactive interface adapted to the MR interaction module (4), and at the same time perform high-fidelity visualization rendering of the virtual content generated by the scene generation submodule (203) through the real-time rendering module.

6. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 5, characterized in that, The MR glasses (401) of the MR interaction module (4) are used to display virtual scenes and collect real-time pose tracking data. The interactive handle (402) is used to receive user operation commands and transmit them to the Unreal Engine scene building module (3) to realize real-time interaction between the user and the MR scene.

7. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 6, characterized in that, The continuous optimization module (5) collects user interaction data fed back by the MR interaction module (4), generates optimization instructions using reinforcement learning algorithms, and transmits them to the artificial intelligence analysis module (2) and the Unreal Engine scene construction module (3) respectively, iteratively optimizing the position, shape, interaction response mechanism and scene visual effects of virtual objects.

8. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 7, characterized in that, The data preprocessing submodule (201) of the artificial intelligence analysis module (2) is used to preprocess the multi-source data transmitted by the multi-sensor fusion module (1), including point cloud denoising, registration, downsampling, as well as image deblurring, enhancement and feature extraction operations.

9. The intelligent MR virtual scene construction system based on Unreal Engine according to claim 8, characterized in that, The Unreal Engine scene building module (3) and the MR interaction module (4) achieve precise spatial alignment between the virtual scene and the real environment through a coordinate transformation algorithm. The coordinate transformation algorithm uses Kalman filtering for pose optimization to improve alignment accuracy.

10. The construction method of an intelligent MR virtual scene construction system based on Unreal Engine as described in any one of claims 1-9, characterized in that, Includes the following steps: Step 1. Start the multi-sensor fusion module (1) and collect three-dimensional point cloud data, RGB image data, depth image data, device pose data and environmental semantic information of the real environment through lidar (101), high-definition camera (102) and inertial measurement unit (103); Step 2. The multi-sensor fusion module (1) transmits the collected multi-source data to the artificial intelligence analysis module (2), and completes the data preprocessing operation through the data preprocessing submodule (201); Step 3. The preprocessed data is semantically parsed by the semantic segmentation submodule (202) to generate a structured environment semantic model containing object semantic labels; Step 4. The environmental semantic model input scene generation submodule (203) automatically generates virtual object models, scene layout schemes and interaction logic through a generative adversarial network model; Step 5. Unreal Engine Scene Building Module (3) receives the above virtual content, builds a basic scene framework in proportion to the real environment, and configures the lighting system, physical collision rules and interaction interface; Step 6. The Unreal Engine scene building module (3) realizes high-fidelity visualization of the virtual scene through the real-time rendering module and establishes communication with the MR interaction module (4); Step 7. The user views the virtual scene through MR glasses (401) and performs interactive operations through the interactive handle (402). The MR interactive module (4) collects real-time pose tracking data and user operation feedback. Step 8. The continuous optimization module (5) receives the data transmitted by the MR interaction module (4) and generates optimization instructions using reinforcement learning algorithms; Step 9. Adjust the generation strategy of the artificial intelligence analysis module (2) and the scene parameters of the Unreal Engine scene construction module (3) according to the optimization instructions to complete the iterative optimization.

Images