A dynamic prism memory scene generation method

Abstract

This invention discloses a dynamic prism memory-based scene generation method in the fields of computer graphics and virtual reality, addressing the problems of existing scene generation technologies, such as lack of effective memory mechanisms and dynamic adaptability, high time cost, large memory consumption, and poor real-time performance in generating complex scenes. It employs an octahedral geometric structure as a logical storage model, mapping eight types of scene data, including geometric, texture, and lighting information, to eight logical storage areas, maintaining data consistency through an 8×8 symmetric association matrix. An intelligent memory navigation matrix is ​​established, using a locality-sensitive hashing algorithm combined with a B+ tree index structure for scene matching. Progressive scene memory fusion is performed, calculating scene differences through a scene differential encoding algorithm and integrating information using an adaptive weight adjustment factor. A hierarchical update strategy is adopted to output the generated scene based on the degree of scene change. Scene generation speed is improved by 52%, and memory consumption is reduced by 26%.

Description

Technical Field

[0001] This invention relates to the fields of computer graphics and virtual reality, and in particular to a scene generation method and system based on dynamic prism memory. Background Technology

[0002] With the rapid development of virtual reality, game development, and digital twin technologies, scene generation technology has become an important research direction in the field of computer graphics. Existing scene generation technologies mainly include procedural generation, rule-based generation, and deep learning-based generation methods.

[0003] Through searching and analysis, several relevant technical solutions were found in the field of scene generation technology. CN116563498A discloses a method for virtual-real fusion in performance and exhibition venues based on the metaverse. This method constructs a 3D model that matches the actual performance scene data, performs keying processing on pre-recorded performance videos, and integrates the 3D model scene and the processed performance video files in a web-based 3D editing engine to achieve a virtual-real fusion effect. However, this solution is mainly aimed at specific performance and exhibition scenarios, lacks universality, and cannot achieve the memory and reuse of scene information.

[0004] CN117274496A discloses a scene construction method based on Unreal Engine. This method creates a virtual scene structure containing a main scene and sub-scenes. The main scene hosts interactive objects and dynamic effects, while the sub-scenes store static meshes. Support objects are generated programmatically and bound to corresponding components. Although this solution adopts a hierarchical scene management structure, it relies on a specific game engine and lacks intelligent scene generation and memory reuse mechanisms.

[0005] CN103885788B discloses a method for constructing dynamic Web 3D virtual reality scenes based on model componentization. This method standardizes and componentizes models according to a unified standard to form a component library, and optimizes the Web 3D scene display by dynamically loading the scene onto a server. Although this solution achieves component-based reuse of models, it is mainly aimed at the Web 3D environment, and the component library has a static structure, lacking dynamic learning and adaptive capabilities.

[0006] The existing technology has the following shortcomings: First, it lacks an effective memory mechanism, making it impossible to store and reuse historically generated scene information, resulting in redundant calculations and wasted resources; second, the generation process lacks dynamic adaptability and cannot be adjusted in real time according to user interaction and environmental changes; third, existing methods have high time costs, large memory consumption, and poor real-time performance in generating complex scenes; and finally, it is difficult to handle the continuous generation of large-scale scenes and lacks a unified representation method for multi-scale scenes. Summary of the Invention

[0007] Technical issues

[0008] The technical problem this invention aims to solve is that existing scene generation technologies lack effective memory mechanisms and dynamic adaptability, resulting in high time costs, large memory consumption, and poor real-time performance in generating complex scenes, making it difficult to achieve intelligent reuse and continuous generation of scene information.

[0009] Technical solution

[0010] To address the aforementioned technical problems, this invention provides a dynamic prism memory scene generation method. This method employs an octahedral geometric structure as the core storage architecture and achieves efficient scene generation through intelligent memory navigation and progressive fusion processing.

[0011] Specifically, this method first constructs an octahedral information prism unit, using an octahedral geometric structure as the logical storage model. Based on its optimal surface area to volume ratio (√2) and the symmetry of its eight triangular faces, it is mapped into a three-dimensional array structure in physical storage. Eight types of scene data—geometric information, texture information, lighting information, semantic information, temporal information, interaction information, material information, and spatial relationship information—are mapped to eight logical storage areas. Geometric information is stored in a three-dimensional coordinate point cloud data format, containing X, Y, and Z coordinate values ​​and normal vector information. Texture information is stored in an RGB color value and UV coordinate mapping format. Lighting information is stored in a format containing light source position, intensity, and color parameters. Each logical storage area uses 32-bit floating-point encoding to store the corresponding type of scene data. Each storage region maintains data consistency through an 8×8 symmetric association matrix. The association matrix is ​​constructed based on the mutual information theory in information theory. The matrix element M[i][j] = I(Xi;Xj) / H(Xi,Xj)×α+e^(-λt)×β, where I(Xi;Xj) is the mutual information between the i-th and j-th types of information, H(Xi,Xj) is the joint entropy, α is the association strength coefficient with a value ranging from 0.1 to 0.9, β is the time decay coefficient with a value ranging from 0.05 to 0.3, λ is the decay constant with a value ranging from 0.001 to 0.01, and t is the time variable.

[0012] The total number of octahedral information prism units is set to 1152–4096, determined according to the number of objects in the scene. Every 100 scene objects correspond to 64–128 storage units, with 128–512 coarse-grained storage units and 1024–3584 fine-grained storage units. The eight geometric faces of the octahedral structure are connected via edge sharing. In physical storage, a three-dimensional array structure is used to form a three-dimensional storage array with spatial indices based on three-dimensional coordinates. The octahedral information prism unit adopts a hierarchical polyhedral storage structure, including coarse-grained storage units storing overall scene layout information and fine-grained storage units storing detailed scene texture information. Hierarchical indexing enables multi-scale scene representation and fast retrieval. The information faces in the octahedral information prism unit are stored in blocks, with each face divided into 16×16 pixel basic storage blocks. Each storage block contains four types of basic data: position coordinates, color values, normal vectors, and material parameters. An input data validation mechanism is set to check the format and range of the scene data.

[0013] This method establishes an intelligent memory navigation matrix, employs Locality Sensitive Hashing (LSH) as the initial screening mechanism, sets the number of hash functions to 8–32, and the number of hash tables to 4–16, and combines it with a B+ tree index structure for precise matching. The first-level index is based on scene type classification, and the second-level index is built based on similarity hash values. The similarity between scene feature vectors is calculated based on Euclidean distance, using the following formula: Where xi is the i-th component of the first scene feature vector and yi is the i-th component of the second scene feature vector, the similarity matching threshold is set to 0.65-0.90. The intelligent memory navigation matrix adopts a hierarchical retrieval strategy. First, preliminary matching is performed at the coarse-grained layer with a matching threshold set to 0.65-0.75. Then, precise matching is performed at the fine-grained layer with a matching threshold set to 0.80-0.90. The index weight is dynamically adjusted according to the usage frequency. The weight update formula is W(t+1)=μW(t)+(1-μ)F(t), where W(t+1) is the weight value at the next time step, W(t) is the weight value at the current time step, μ is the decay coefficient of 0.8-0.9, and F(t) is the current usage frequency.

[0014] This method performs progressive scene memory fusion. It calculates the spatial structure differences and content changes between the new scene and the reference scene using a scene differential coding algorithm. A weighted fusion approach is then used to integrate the scene information. The initial range of the fusion weight coefficients is set to 0.3–0.7, and these coefficients are adaptively adjusted based on the scene complexity C = 0.4 × (number of polygons / 10000) + 0.3 × (texture resolution / 1024²) + 0.3 × (number of light sources / 10). The scene differential coding algorithm in progressive scene memory fusion uses a spatial structure comparison method, dividing the scene into an N×N grid. It calculates the geometric change rate G_diff = Σ|P_new(i) - P_ref(i)| / N_points, the texture difference T_diff = Σ|RGB_new(i) - RGB_ref(i)| / (3×N_pixels), and the illumination change amplitude L_diff = |I_new - I_ref| / I_max, where P_new(i) is... The position coordinates of the i-th point in the new scene, P_ref(i) are the position coordinates of the i-th point in the reference scene, N_points are the total number of points, RGB_new(i) are the RGB values ​​of the i-th pixel in the new scene, RGB_ref(i) are the RGB values ​​of the i-th pixel in the reference scene, N_pixels are the total number of pixels, I_new is the illumination intensity of the new scene, I_ref is the illumination intensity of the reference scene, and I_max is the maximum illumination intensity. The weight coefficients of the weighted fusion are adaptively adjusted according to the scene complexity.

[0015] This method outputs a generated scene, generates a target scene based on the fused scene information, and updates the memory information in the octahedral information prism unit using a hierarchical update strategy according to the degree of scene change. In the hierarchical update strategy, scene regions with pixel differences less than 2% are updated at a low frequency of 30 times / second, scene regions with pixel differences between 2% and 8% are updated at a medium frequency of 60 times / second, and scene regions with pixel differences greater than 8% are updated at a high frequency of 90 times / second.

[0016] This invention also provides a scene generation system for implementing the above method, including an octahedral memory management module, an intelligent navigation index module, a scene fusion processing module, and a scene generation output module. The octahedral memory management module is used to construct and manage the octahedral information prism unit, including a GPU memory cache unit, a memory management unit, and a dimension data cache. The GPU memory cache unit supports 4GB to 32GB of video memory configuration, adopts a hierarchical caching strategy, with a limit of 1152 storage units for 4GB configuration and a complete 8×8 association matrix for configurations above 8GB. The memory management unit is responsible for the creation, updating, and deletion of storage units. The dimension data cache includes an association matrix calculation unit and a data synchronization controller. The octahedral memory management module also includes a memory compression unit, which uses a wavelet transform-based compression algorithm to compress the storage units, achieving a compression ratio of 2:1 to 3:1. The memory management unit uses an LRU cache replacement strategy to manage the lifecycle of the storage units. The intelligent navigation index module is used to implement the intelligent memory navigation matrix, supporting parallel retrieval with CUDA computing power of 7.0 or higher. The scene fusion processing module is used to execute the progressive scene memory fusion. The scene generation output module is used to output the generated scene and update the memory information.

[0017] The system supports distributed deployment, including a cloud memory server and edge computing nodes. The cloud memory server stores a global storage unit library, and the edge computing nodes cache commonly used storage units. Data is synchronized through a 5G network, and AES-256 encryption algorithm is used to protect data transmission, achieving end-to-end encryption. Data transmission latency is controlled within the range of 50 to 200 ms.

[0018] Beneficial effects

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] 1. Significantly improved scene generation efficiency: Through the octahedral information prism memory mechanism, intelligent reuse of scene information is achieved, increasing scene generation speed by 52% and reducing the generation time of a single scene to 1.2 seconds, meeting the needs of real-time applications.

[0021] 2. Significantly improved memory efficiency: Adopting an octahedral information prism structure and intelligent memory navigation matrix, memory usage is reduced by 26%, while supporting dynamic memory management to avoid memory fragmentation issues.

[0022] 3. Strong dynamic adaptability: It has adaptive learning and parameter adjustment capabilities, and can dynamically optimize the generation strategy according to changes in usage patterns and scenarios, achieving a similarity matching accuracy of 89%.

[0023] 4. Unified representation of multi-scale scenes: Supports multi-level scene representation from macro layout to micro details, enabling efficient generation of large-scale continuous scenes and improving scene consistency by more than 70%. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall architecture of the dynamic prism memory scene generation system according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the octahedral information prism unit structure according to an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the scene generation method according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0028] Example 1

[0029] The purpose of this embodiment is to verify the complete technical solution of the scene generation method based on the octahedral information prism memory unit and to demonstrate the technical effect of the dynamic memory mechanism in improving scene generation efficiency.

[0030] like Figure 1 As shown, the dynamic prism memory scene generation system constructed in this embodiment includes an octahedral memory management module, an intelligent navigation index module, a scene fusion processing module, and a scene generation output module.

[0031] like Figure 3 As shown, the scene generation method of the present invention includes four main steps: constructing an octahedral information prism unit, establishing an intelligent memory navigation matrix, performing progressive scene memory fusion, and outputting the generated scene.

[0032] Step 1: Construct an octahedral information prism unit. An octahedral geometric structure is used as the logical storage model, such as... Figure 2 As shown, the octahedron has eight triangular faces, each corresponding to a type of scene information: the geometric information face stores 3D coordinate point cloud data, including X, Y, and Z coordinate values ​​(accuracy 0.001m) and normal vector information (unit vector); the texture information face stores RGB color values ​​(range 0-255) and UV coordinate mapping (range 0.0-1.0); the lighting information face stores the light source position coordinates, light intensity value (0-1000 lux), and light source color parameters; the semantic information face stores object category labels and attribute descriptions; the temporal information face stores scene timestamps and change frequencies; the interaction information face stores user interaction history and response parameters; the material information face stores material type, reflectivity, and transparency parameters; and the spatial relationship information face stores the positional relationships and constraints between objects.

[0033] Each logical storage area uses 32-bit floating-point encoding, with a storage capacity of 4KB, totaling 32KB / unit. An 8×8 symmetric association matrix M is established, constructed based on the mutual information theory in information theory. The matrix element M[i][j] = I(Xi;Xj) / H(Xi,Xj)×α + e^(-λt)×β, where I(Xi;Xj) is the mutual information between the i-th and j-th types of information, H(Xi,Xj) is the joint entropy, α takes a value of 0.6 (association strength coefficient), β takes a value of 0.15 (time decay coefficient), λ takes a value of 0.005 (decay constant), and t is the time variable. The time decay factor adopts an exponential decay model, conforming to the memory forgetting curve. In this embodiment, a total of 2048 octahedral information prism units are set, including 256 coarse-grained storage units and 1792 fine-grained storage units.

[0034] Step Two: Establish an Intelligent Memory Navigation Matrix. Locality Sensitive Hash (LSH) is used as the initial screening mechanism, with 16 hash functions and 8 hash tables. A B+ tree index structure is combined for precise matching; the B+ tree is of order 128, and the leaf nodes store index pointers to scene feature vectors. The first-level index is based on scene type classification, including 8 main categories such as indoor scenes, outdoor scenes, natural scenes, and urban scenes; the second-level index is built based on similarity hash values, with each hash value being 64 bits long.

[0035] Scene feature vector extraction employs a multi-dimensional feature fusion method, including geometric features (bounding boxes, volume, surface area), texture features (color histogram, texture orientation gradient), lighting features (light intensity distribution, shadow coverage), and semantic features (object category distribution, spatial layout pattern). The similarity between scene feature vectors is calculated based on Euclidean distance, using the following formula: , where xi is the i-th component of the first scene feature vector and yi is the i-th component of the second scene feature vector. The similarity matching threshold is set to 0.85; a scene match is considered successful when the similarity is greater than the threshold.

[0036] Step 3: Perform progressive scene memory fusion. The scene differential coding algorithm uses a 16×16 grid block comparison method to calculate the geometric change rate, texture difference, and illumination change amplitude between the new scene and the reference scene. The geometric change rate G_diff = Σ|P_new(i) - P_ref(i)| / N_points, where P_new(i) is the position coordinate of the i-th point in the new scene, P_ref(i) is the position coordinate of the i-th point in the reference scene, and N_points is the number of point clouds; the texture difference T_diff = Σ|RGB_new(i) - RGB_ref(i)| / (3×N_pixels), where RGB_new(i) is the RGB value of the i-th pixel in the new scene, RGB_ref(i) is the RGB value of the i-th pixel in the reference scene, and N_pixels is the number of pixels; the illumination change amplitude L_diff = |I_new - I_ref| / I_max, where I_new is the illumination intensity of the new scene, I_ref is the illumination intensity of the reference scene, and I_max is the maximum illumination intensity.

[0037] An adaptive weighting factor is used for scene information integration. The initial value of the fusion weight coefficient is set to 0.5 and adjusted according to the scene complexity. Scene complexity is calculated using the formula C = 0.4 × (number of polygons / 10000) + 0.3 × (texture resolution / 1024²) + 0.3 × (number of light sources / 10). When the complexity is less than 0.3, the weight coefficient is adjusted to 0.3; when the complexity is greater than 0.7, the weight coefficient is adjusted to 0.7; otherwise, the weight coefficient equals the complexity value.

[0038] Step 4: Output the generated scene. Based on the fused scene information, a target scene is generated using a tiered update strategy: low-frequency updates (30 times / second) are used for scene areas with pixel differences less than 2%; medium-frequency updates (60 times / second) are used for scene areas with pixel differences between 2% and 8%; and high-frequency updates (90 times / second) are used for scene areas with pixel differences greater than 8%.

[0039] Test environment: NVIDIA RTX 4090 graphics card (24GB VRAM), Intel i9-13900K processor, 32GB DDR5 memory, Windows 11 operating system. The test dataset contains 1000 different types of 3D scenes, ranging in complexity from simple indoor scenes to complex urban scenes.

[0040] This embodiment verifies that the average scene generation time is 1.2 seconds, a 52% improvement compared to the traditional method's 2.5 seconds; the average memory usage is 3.1GB, a 26% saving compared to the traditional method's 4.2GB; the real-time rendering frame rate reaches 62 FPS, exceeding the real-time rendering requirements; the Scene Quality Score (SSIM) reaches 0.89, and the generation success rate is 98.5%. This demonstrates the significant advantages of the scene generation method based on octahedral information prism memory in improving generation efficiency and reducing resource consumption.

[0041] Example 2

[0042] The purpose of this embodiment is to verify the technical effect of the boundary value of the number of storage units and to demonstrate the impact of different storage unit configurations on system performance.

[0043] Based on Example 1, the performance was tested with 1152, 2048, and 4096 storage units respectively. Other parameters remained the same as in Example 1.

[0044] The test results are as follows:

[0045] 1152 storage units: scene generation time 1.4 seconds, memory usage 2.8GB, cache hit rate 76%, suitable for medium complexity scenes;

[0046] With 2048 storage units, the scene generation time is 1.2 seconds, memory usage is 3.1GB, cache hit rate is 85%, and the performance is optimal.

[0047] With 4096 storage units, the scene generation time is 1.0 second, the memory usage is 4.2GB, and the cache hit rate is 92%, making it suitable for highly complex scenarios.

[0048] This embodiment verifies that a storage unit count of 1152-4096 can adapt to the generation requirements of scenarios with different complexities, and 2048 storage units is the optimal configuration, achieving the best balance between performance and resource consumption.

[0049] Example 3

[0050] The purpose of this embodiment is to verify the technical effect of the similarity matching threshold boundary value and to demonstrate the influence of the threshold parameter on system performance.

[0051] Based on Example 1, the performance was tested with similarity matching thresholds of 0.65, 0.75, 0.85, and 0.90, respectively. Other parameters remained the same as in Example 1.

[0052] The test results are as follows:

[0053] Threshold 0.65: Scene generation time 1.0 second, memory usage 2.8GB, matching accuracy 78%, false match rate 15%;

[0054] Threshold 0.75: Scene generation time 1.1 seconds, memory usage 2.9GB, matching accuracy 83%, false match rate 8%;

[0055] Threshold 0.85: Scene generation time 1.2 seconds, memory usage 3.1GB, matching accuracy 89%, false match rate 5%;

[0056] Threshold 0.90: Scene generation time 1.4 seconds, memory usage 3.3GB, matching accuracy 92%, false match rate 3%;

[0057] This embodiment verifies that a similarity matching threshold in the range of 0.65-0.90 can achieve a good balance between matching accuracy and generation efficiency. A threshold of 0.85 is the optimal choice, which ensures both high matching accuracy and good generation efficiency.

[0058] Example 4

[0059] The purpose of this embodiment is to verify the technical effect of the boundary value of the fusion weight coefficient and to demonstrate the impact of the weight parameter on the quality of scene fusion.

[0060] Based on Example 1, the scene fusion effects were tested with fusion weight coefficients of 0.3, 0.5, and 0.7, respectively. The test scene was a medium-complexity indoor scene with a complexity value of 0.45.

[0061] The test results are as follows:

[0062] Weighting coefficient 0.3: Scene fusion time 0.8 seconds, fusion quality score 0.82, new information retention rate 65%, historical information retention rate 85%;

[0063] Weighting coefficient 0.5: Scene fusion time 0.9 seconds, fusion quality score 0.89, new information retention rate 75%, historical information retention rate 75%;

[0064] Weighting coefficient 0.7: Scene fusion time 1.0 second, fusion quality score 0.86, new information retention rate 85%, historical information retention rate 65%;

[0065] This embodiment verifies that a fusion weighting coefficient of 0.5 can achieve the best balance between new and historical information, obtaining the highest fusion quality score, thus proving the rationality of the parameter range of 0.3-0.7.

[0066] Example 5

[0067] The purpose of this embodiment is to verify the system adaptability under different hardware configurations and demonstrate the technical effect of the hierarchical processing strategy.

[0068] Tests were conducted on three configurations: RTX 3060 (8GB VRAM), RTX 3080 (10GB VRAM), and RTX 4090 (24GB VRAM).

[0069] RTX 3060 configuration: storage unit count limited to 1152, using a simplified 4×4 association matrix, scene generation time 1.8 seconds, memory usage 2.5GB;

[0070] RTX 3080 configuration: 2048 storage units, using a complete 8×8 association matrix, scene generation time 1.4 seconds, memory usage 3.2GB;

[0071] RTX 4090 configuration: 4096 storage units, using a complete 8×8 association matrix, scene generation time 1.0 second, memory usage 4.2GB;

[0072] This embodiment verifies the system's good adaptability under different hardware configurations, and the hierarchical processing strategy effectively reduces the dependence on high-end hardware.

[0073] Comparative Example 1

[0074] The purpose of this comparative example is to verify the difference in technical performance when the octahedral information prism memory unit is missing, and to highlight the key role of the dynamic memory mechanism through comparison.

[0075] The contrasting approach employs a traditional static caching mechanism, removing the octahedral information prism memory unit and replacing it with a linear array structure to store scene information. Scene information is stored in separate data structures according to categories such as geometry, texture, and lighting, lacking a unified multi-dimensional representation and association mechanism. The retrieval algorithm uses linear search, with a time complexity of O(n).

[0076] The test conditions were the same as in Example 1, using the same hardware environment and test dataset. Test results showed that: scene generation time was 3.2 seconds, an increase of 167% compared to 1.2 seconds in Example 1; memory usage was 4.8GB, an increase of 55% compared to 3.1GB in Example 1; cache hit rate was only 28%, far lower than 85% in Example 1; and the Scene Quality Score (SSIM) was 0.76, lower than 0.89 in Example 1.

[0077] The comparison shows that the lack of octahedral information prism memory units leads to a significant decrease in scene generation efficiency, reduced memory usage efficiency, and a decline in scene quality, demonstrating the crucial role of octahedral information prism memory units in improving the overall system performance.

[0078] Comparative Example 2

[0079] The purpose of this comparative example is to verify the difference in technical effectiveness when an 8×8 symmetric correlation matrix is ​​missing.

[0080] The comparative scheme removes the 8×8 symmetric correlation matrix, and each information surface is stored and processed independently, without performing correlation calculations or consistency maintenance. Other technical features remain the same as in Example 1.

[0081] Test results show that: the scene generation time is 1.8 seconds, which is 50% longer than that of Example 1; the data consistency error rate is 12%, compared to 2% in Example 1; the scene fusion quality score is 0.81, which is lower than 0.89 in Example 1; the system stability has decreased, and the probability of data conflict is 8%.

[0082] The comparative analysis verified the important role of the 8×8 symmetric correlation matrix in maintaining the consistency of multidimensional information and improving the fusion quality.

[0083] Comparative Example 3

[0084] The purpose of this comparison is to verify the performance differences with existing scene generation technologies for Unity and Unreal Engine.

[0085] The comparison schemes adopted the standard scene generation workflows of Unity 2022.3 and Unreal Engine 5.1, respectively, and used the same test dataset and hardware environment.

[0086] The test results are compared below:

[0087]

[0088] Comparative analysis has verified the significant advantages of this invention over existing mainstream engines in terms of scene generation efficiency, memory usage, and quality.

[0089] Based on the above embodiments and comparative examples, the core innovation of this invention lies in constructing a dynamic memory mechanism using octahedral information prism memory units. An 8×8 symmetric association matrix is ​​used to achieve unified representation and consistency maintenance of multi-dimensional information. Combined with an intelligent memory navigation matrix and progressive scene fusion processing, efficient scene generation and intelligent reuse are achieved. Compared with existing technologies, this invention has significant advantages in scene generation efficiency, memory utilization efficiency, and dynamic adaptability, providing a high-performance scene generation solution for virtual reality and digital twin applications.

Claims

1. A method for generating scenes based on dynamic prism memory, characterized in that, Includes the following steps: Step 1: Construct an octahedral information prism unit. This unit uses an octahedral geometric structure to organize its internal logical storage model. By mapping eight types of scene data to the logical storage areas corresponding to the eight geometric faces of the octahedron, the data is structured. Its physical storage form is implemented using a three-dimensional array structure. The eight types of scene data—geometric information, texture information, lighting information, semantic information, temporal information, interaction information, material information, and spatial relationship information—are mapped to eight logical storage areas. Geometric information is stored in a three-dimensional coordinate point cloud data format, containing X, Y, and Z coordinate values ​​and normal vector information. Texture information is stored in an RGB color value and UV coordinate mapping format. Lighting information is stored in a light source position, intensity, and color parameter format. Each logical storage area uses 32-bit floating-point encoding to store the corresponding type of scene data. Data transmission and consistency maintenance are achieved between the logical storage areas through an 8×8 symmetric association matrix, ensuring the correlation and synchronization between different types of scene data. The matrix is ​​constructed based on the mutual information theory in information theory. The matrix element M[i][j] = I(Xi;Xj) / H(Xi,Xj)×α + e^(-λt)×β, where I(Xi;Xj) is the mutual information between the i-th and j-th classes of information, H(Xi,Xj) is the joint entropy, α is the correlation strength coefficient (ranging from 0.1 to 0.9), β is the time decay coefficient (ranging from 0.05 to 0.3), λ is the decay constant (ranging from 0.001 to 0.01), and t is the time constant. Interval variables; the total number of the octahedral information prism units is set to 1152-4096, which is determined according to the number of objects contained in the scene. Every 100 scene objects correspond to 64-128 storage units, of which the number of coarse-grained storage units is 128-512 and the number of fine-grained storage units is 1024-3584. The eight geometric faces of the octahedral structure are connected by edge sharing and are physically stored using a three-dimensional array structure to form a three-dimensional storage array with spatial index as three-dimensional coordinates; Step Two: Establish an intelligent memory navigation matrix, employing Locality Sensitive Hash (LSH) as the initial screening mechanism. The number of hash functions is set to 8–32, and the number of hash tables is set to 4–16. A B+ tree index structure is used for precise matching. The first-level index is based on scene type classification, and the second-level index is built based on similarity hash values. The similarity between scene feature vectors is calculated using Euclidean distance, with the following formula: , where xi is the i-th component of the first scene feature vector, yi is the i-th component of the second scene feature vector, and the similarity matching threshold is set to 0.65~0.90; Step 3: Perform progressive scene memory fusion. Calculate the spatial structure differences and content changes between the new scene and the reference scene using a scene differential coding algorithm. Integrate the scene information using a weighted fusion method. The initial range of the fusion weight coefficient is set to 0.3–0.

7. The fusion weight coefficient is adaptively adjusted based on the scene complexity C = 0.4 × (number of polygons / 10000) + 0.3 × (texture resolution / 1024²) + 0.3 × (number of light sources / 10). Step 4: Output the generated scene. Generate the target scene based on the fused scene information, and adopt a hierarchical update strategy according to the degree of scene change. Update the memory information in the octahedral information prism unit through the data transmission mechanism of the association matrix.

2. The method for generating a scene based on a dynamic prism memory according to claim 1, characterized in that, The intelligent memory navigation matrix adopts a hierarchical retrieval strategy. First, preliminary matching is performed at the coarse-grained layer with a matching threshold set to 0.65–0.

75. Then, precise matching is performed at the fine-grained layer with a matching threshold set to 0.80–0.

90. The index weight is dynamically adjusted according to the usage frequency. The weight update formula is W(t+1) = μW(t) + (1-μ)F(t), where W(t+1) is the weight value at the next time step, W(t) is the weight value at the current time step, μ is the attenuation coefficient of 0.8–0.9, and F(t) is the current usage frequency.

3. The method for generating a scene based on a dynamic prism memory according to claim 1, characterized in that, The scene differential encoding algorithm in the progressive scene memory fusion adopts a spatial structure comparison method, dividing the scene into an N×N grid, and calculating the geometric change rate G_diff=Σ|P_new(i)-P_ref(i)| / N_points, texture difference T_diff=Σ|RGB_new(i)-RGB_ref(i)| / (3×N_pixels), and illumination change amplitude L_diff=|I_new-I_ref| / I_max, where P_new(i) Let P_ref(i) be the position coordinate of the i-th point in the new scene, P_ref(i) be the position coordinate of the i-th point in the reference scene, N_points be the total number of points, RGB_new(i) be the RGB value of the i-th pixel in the new scene, RGB_ref(i) be the RGB value of the i-th pixel in the reference scene, N_pixels be the total number of pixels, I_new be the illumination intensity of the new scene, I_ref be the illumination intensity of the reference scene, and I_max be the maximum illumination intensity. The weight coefficients of the weighted fusion are adaptively adjusted according to the scene complexity.

4. The method for generating a scene based on a dynamic prism memory according to claim 1, characterized in that, In the tiered update strategy, low-frequency updates of 30 times / second are used for scene areas with pixel differences of less than 2%, medium-frequency updates of 60 times / second are used for scene areas with pixel differences between 2% and 8%, and high-frequency updates of 90 times / second are used for scene areas with pixel differences of more than 8%.

5. The method for generating a scene based on a dynamic prism memory according to claim 1, characterized in that, The information facets in the octahedral information prism unit are stored in blocks. Each information facet is divided into a 16×16 pixel basic storage block. Each storage block contains four types of basic data: position coordinates, color value, normal vector, and material parameters. An input data verification mechanism is set to check the format and range of the scene data.

6. The method for generating a scene based on a dynamic prism memory according to claim 1, characterized in that, The octahedral information prism unit adopts a hierarchical polyhedral storage structure, including a coarse-grained storage unit for storing the overall layout information of the scene and a fine-grained storage unit for storing the detailed texture information of the scene. It achieves multi-scale scene representation and fast retrieval through hierarchical indexing.

7. A scene generation system, characterized in that, To implement the dynamic prism memory scene generation method of claim 1, the system comprises: an octahedral memory management module for constructing and managing the octahedral information prism unit, including a GPU memory cache unit, a memory management unit, and a dimension data cache. The GPU memory cache unit supports 4GB to 32GB of video memory configuration and adopts a hierarchical caching strategy. The number of storage units is limited to 1152 under the 4GB video memory configuration, and a complete 8×8 association matrix is ​​used for configurations above 8GB. The memory management unit is responsible for the creation, updating, and deletion of storage units. The dimension data cache includes an association matrix calculation unit and a data synchronization controller; an intelligent navigation index module for implementing the intelligent memory navigation matrix and supporting parallel retrieval with CUDA computing power of 7.0 or higher; a scene fusion processing module for performing the progressive scene memory fusion; and a scene generation output module for outputting the generated scene and updating the memory information.

8. The scene generation system according to claim 7, characterized in that, The octahedral memory management module also includes a memory compression unit, which uses a wavelet transform-based compression algorithm to compress the storage units, achieving a compression ratio of 2:1 to 3:

1. The memory management unit uses an LRU cache replacement strategy to manage the lifecycle of the storage units.

9. The scene generation system according to claim 7, characterized in that, The system supports distributed deployment, including a cloud memory server and edge computing nodes. The cloud memory server stores a global storage unit library, and the edge computing nodes cache commonly used storage units. Data is synchronized through a 5G network, and AES-256 encryption algorithm is used to protect data transmission, achieving end-to-end encryption. Data transmission latency is controlled within the range of 50 to 200 ms.

Images