A VR large space rendering method and system based on a cloud server cluster

Abstract

The application discloses a VR large-space rendering method and system based on a cloud server cluster, which comprises the following steps: partitioning large-scale complex three-dimensional data and constructing a multi-scale level of detail (LOD) based on domain prior knowledge; performing analysis-intention-driven rendering and predictive data loading based on a domain model; and applying a priority synchronization strategy for collaborative interaction. The application solves the problem of real-time interaction and collaborative analysis of large-scale complex three-dimensional data in a VR environment by deeply integrating domain knowledge and cloud rendering technology, and improves the rendering efficiency, interaction fluency and logical consistency of multi-user collaboration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of virtual reality (VR), cloud computing and computer graphics, in particular to a VR large space rendering method and system based on a cloud server cluster. BACKGROUND

[0002] With the development of three-dimensional scanning, scientific computing and big data technology, unprecedentedly large and complex PB (Petabyte) level three-dimensional data sets have been generated in many fields such as city models, industrial design, geological exploration and bioinformatics. Immersive and interactive visualization exploration of these large-scale data in a virtual reality (VR) environment has important value for scientific discovery, engineering decision-making and education and training. However, such a large amount of data poses a serious challenge to existing computing and rendering technology.

[0003] In the prior art, there are limitations in VR rendering of large-scale three-dimensional data. On the one hand, traditional single-machine or desktop-level VR applications are limited by the computing, storage and display memory capabilities of local hardware, and cannot directly load and render PB-level data in real time. Users can only observe a low-precision overview or a small subset of the data after segmentation, and cannot obtain a global understanding and a smooth interactive experience. On the other hand, although existing cloud rendering or pixel streaming technology can push the rendering results of a remote server to a lightweight VR client in the form of a video stream, solving the hardware bottleneck of the client, its rendering strategy and data scheduling mechanism are usually general. For example, its data loading and prefetching strategy is mostly based on the user's current viewpoint position or the general principle of spatial proximity. For large space scenes with non-uniform data distribution and complex structure, such as neural network data with complex topological connections or digital city models with rich geometric details, traditional strategies are inefficient and cannot accurately allocate valuable computing and bandwidth resources to the target areas of real interest to the user, resulting in high interaction delay, frame detail switching lag and other problems. SUMMARY

[0004] The technical problem to be solved by the present application is that the prior art has low visualization interaction efficiency, lacks perception of user analysis intent, and cannot guarantee consistency of collaborative analysis when processing large-scale complex three-dimensional data.

[0005] To solve the above technical problems, one aspect of the present application provides a VR large space rendering method based on a cloud server cluster. The method organizes data at multiple scales by combining domain prior knowledge, and uses an analysis intent-driven and predictive loading rendering strategy, as well as a priority synchronization mechanism for collaborative interaction, to improve the efficiency and experience of visualization analysis of large-scale complex three-dimensional data. The method includes the following steps:

[0006] Step one involves partitioning and constructing a multi-scale level of detail (LOD) structure for the large-scale complex 3D data based on domain prior knowledge. This domain prior knowledge refers to existing structured knowledge within a specific domain, such as anatomical maps of the human brain in neuroscience, or administrative divisions and road network topology in the field of digital cities. This step reorganizes the raw, massive data into a hierarchical structure suitable for efficient streaming scheduling and rendering in the cloud. Specifically, this step includes:

[0007] Sub-step 1: Use domain-specific spatial partitioning criteria to perform macroscopic pre-partitioning of the large-scale complex 3D data, initially dividing the data into several macroscopic data blocks with clear domain significance.

[0008] Sub-step 2 involves using community detection or clustering algorithms based on domain characteristics to identify and optimize functional clusters within each macro-partition, thereby forming more refined data organization units.

[0009] Sub-step 3: Based on the above partitioning results, construct a multi-scale level of detail (LOD) model. This model abstracts the connection of complex data units at lower levels of detail, and resolves them into high-resolution, independent data units or their connections at higher levels of detail, thereby achieving efficient rendering of data at different observation scales.

[0010] Step two involves executing an analysis intent-driven rendering and predictive data loading. This step intelligently schedules and renders data based on the user's exploration tasks. Specifically, when the rendering engine deployed on the cloud server cluster receives a user analysis command from the virtual reality client, it parses the semantics of the command, identifies the key data partitions and functional clusters involved, and streams and renders the associated high-LOD level data with high priority. It also initiates a domain model-based predictive data loading mechanism. This mechanism utilizes a preset model that describes the interaction or state transition patterns between data units. Based on the state of the data clusters currently observed or activated by the user, it performs forward-looking behavior predictions, calculates the downstream data clusters most likely to be affected or activated within a preset time window, and asynchronously preloads the high-LOD level data associated with the downstream data clusters into the rendering cache.

[0011] Step 3: Apply a priority synchronization strategy for collaborative interaction. This step ensures logical consistency of key analytical operations and accuracy of data synchronization when multiple users are collaborating in the same virtual reality space. Specifically, this step includes:

[0012] Sub-step 1: Establish a priority sorting mechanism that is semantically bound to the interaction operation type. Classify and prioritize the operation instructions issued by the user in the collaborative environment according to their domain semantics. For example, they can be divided into analytical operations with higher priority, annotation operations with medium priority, and navigation operations with lower priority.

[0013] Sub-step 2: Based on the priority of the operation, a differential synchronization mechanism is used to synchronize data. High-priority operations are synchronized through a high-reliability data channel to ensure the logical consistency of critical operations, while low-priority operations can be synchronized through a non-reliable channel that allows for packet loss to save network bandwidth.

[0014] Another aspect of the present invention provides a VR large-space rendering system based on a cloud server cluster, the system comprising:

[0015] The data preprocessing module is configured to perform the data partitioning and multi-scale detail level construction based on domain prior knowledge as described in step one of the above methods;

[0016] The cloud rendering engine module, deployed on a cloud server cluster, is configured to perform the analysis intent-driven rendering and predictive data loading described in step two of the above method.

[0017] The collaborative synchronization server module is configured to execute the application-oriented collaborative interaction priority synchronization strategy described in step three of the above method.

[0018] And one or more virtual reality clients, configured to present rendering results to users and receive user interaction commands to send to the cloud rendering engine module and the collaborative synchronization server module.

[0019] In summary, this invention discloses a VR large-space rendering method and system based on a cloud server cluster, comprising: partitioning and constructing multi-scale levels of detail (LOD) of large-scale complex 3D data based on domain prior knowledge; performing analysis intent-driven rendering and predictive data loading based on the domain model; and applying a priority synchronization strategy for collaborative interaction. This invention, by deeply integrating domain knowledge with cloud rendering technology, solves the problem of real-time interaction and collaborative analysis of large-scale complex 3D data in a VR environment, improving rendering efficiency, interaction smoothness, and logical consistency of multi-user collaboration. Compared with existing technologies, the beneficial effects of this invention are:

[0020] 1. By adopting a data partitioning and multi-scale LOD construction method based on domain prior knowledge, this invention reorganizes massive, disordered, large-scale complex 3D data into a hierarchical structure with domain significance. Compared with existing solutions that use general segmentation and geometric simplification techniques, this improves the efficiency of subsequent data querying and streaming scheduling, and enhances the domain interpretability of visualization results.

[0021] 2. By introducing an analysis intent-driven rendering mechanism and further combining it with predictive data loading based on domain models, this invention can intelligently and proactively focus computing and rendering resources on the areas most relevant to the user's exploration task. This overcomes the blindness of traditional viewpoint-based rendering strategies when dealing with data with complex internal structures, thereby achieving smooth, low-latency real-time interactive exploration under massive data scales.

[0022] 3. By designing a priority synchronization strategy for collaborative interaction, this invention provides reliable technical support for multi-user collaborative analysis scenarios, ensuring the logical consistency and accuracy of results for critical and global analysis operations. Compared with the synchronization mechanism of general VR collaborative platforms, it is more suitable for collaborative work scenarios in professional fields. Attached Figure Description

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart illustrating a VR large-space rendering method based on a cloud server cluster, provided by an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the principle of data partitioning and multi-scale level of detail (LOD) construction based on domain prior knowledge in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram illustrating the working principle of the user intent-driven rendering and predictive data loading mechanism in this embodiment of the invention.

[0027] Figure 4 This is a schematic diagram illustrating the principle of a priority synchronization strategy for collaborative interaction in an embodiment of the present invention.

[0028] Figure 5 This is a structural block diagram of a VR large-space rendering system based on a cloud server cluster provided in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram showing the hierarchical comparison of the multi-scale level of detail (LOD) model in an embodiment of the present invention.

[0030] Figure 7 This is a performance comparison test chart between the embodiments of the present invention and the prior art. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0032] Example 1

[0033] This embodiment provides a VR large-space rendering method based on a cloud server cluster. Please refer to [link / reference]. Figure 1 The method mainly includes the following steps:

[0034] Step S1: Based on domain prior knowledge, partition the large-scale complex 3D data and construct a multi-scale level of detail (LOD).

[0035] This step reorganizes the raw, unstructured massive amounts of data into a hierarchical, domain-specific data structure that can be efficiently streamed and rendered by a cloud rendering engine.

[0036] The following section uses the processing of petabyte-scale whole-brain connectome data as an example to explain this step in detail:

[0037] Step S101: Perform macroscopic pre-partitioning of large-scale complex 3D data using domain-specific spatial partitioning standards.

[0038] In one specific implementation, taking neural connectome data as an example, the raw data can be represented as a graph structure G=(V, E), where V represents the set of neurons and E represents the set of synaptic connections between them. This step selects a recognized brain atlas, such as the AAL (Automated Anatomical Labeling) atlas containing 90 anatomical regions or the Broadman partitioning atlas, as its domain-specific spatial partitioning standard. By mapping the three-dimensional spatial coordinates of each neuron to the corresponding anatomical region in the atlas, a region label R(v) is assigned to each neuron vertex v∈V. In this way, the entire graph data G is initially divided into several macroscopic data blocks, each corresponding to a specific brain functional area. For example, all neurons located in the "hippocampus" region and their internal connections are grouped into one data block. This pre-partitioning operation allows subsequent data queries to be based directly on meaningful region names, rather than complex spatial coordinate indexes, improving data retrieval efficiency.

[0039] Step S102: Within each macro-partition, functional clusters are identified and optimized using community detection or clustering algorithms based on domain characteristics weighting.

[0040] Taking neural connectome data as an example, after completing the macro-level partitioning, this step employs an improved community detection algorithm to further explore the fine structure within each partition. Specifically, the Louvain community detection algorithm can be used, but the weights for its modularity optimization are redefined. For any connection between two data units (neurons) i and j, the weight w(i,j) is a weighted sum of multiple domain characteristics (biological characteristics), and its calculation formula can be: w(i,j) = α * S(i,j) + β * T(i,j) + γ / D(i,j). Here, S(i,j) is the standardized connection strength (e.g., synaptic connection strength), which can be standardized by linearly mapping the original value to the interval [0, 1]; T(i,j) is the data unit type similarity function (e.g., a value of 1 when two neurons are both excitatory or inhibitory, and a value of 0 otherwise); D(i,j) is the spatial distance between them (e.g., Euclidean distance); α, β, γ are preset weight coefficients, which sum to 1, thus constructing a weighted graph with neurons as nodes and the weights as connection edge weights. For example, in a preferred configuration, α=0.6, β=0.3, γ=0.1. By running a community detection algorithm on this weighted graph, highly functionally related and tightly connected clusters of data units can be identified within each macro-partition. By adopting this domain-specific weighted clustering method, the data organization structure is more consistent with the functional organization of the real world. Compared with general segmentation algorithms that only consider topological connections, the data organization itself contains domain insights, laying the foundation for subsequent analysis intent-driven rendering.

[0041] Step S103: Based on the above partitioning results, construct a multi-scale level of detail (LOD) model.

[0042] For example, regarding connectome data, please refer to [link / reference]. Figure 2 and Figure 6 , Figure 2 This illustrates the basic principles of the LOD model. Figure 6 A detailed comparison of the three key levels is presented. At the lowest level of detail (LOD 0), when a user observes the entire dataset in VR, the system only renders abstract geometry representing the various macroscopic partitions, such as... Figure 6 The left side shows a macroscopic partitioning abstraction of areas A to E. When the user zooms in to the medium level of detail (LOD 1), the system begins rendering the functional clusters identified in step S102, as well as the hyperedges connecting different clusters, such as... Figure 6The functional clusters and hyperedge connections shown in the middle are graphical representations of a large number of real data connections. When the user focuses on one or more clusters, the system seamlessly switches to the highest level of detail (LOD 2). At this point, the hyperedge is resolved into high-resolution, individually identifiable, and interactive single data connections or data units, such as... Figure 6 The high-resolution cell connection network is shown on the right. This LOD model can be implemented by managing a spatial index structure (such as an octree). During rendering, the appropriate LOD level is dynamically selected and rendered based on the distance between the viewpoint and the data block. This method effectively reduces the amount of data to be rendered per frame while maintaining visual continuity.

[0043] Step S2 involves performing an analysis intent-driven rendering and predictive data loading.

[0044] When users interact in the VR environment, this step prioritizes allocating limited computing and network resources to the data areas that users are most interested in, thereby achieving a smooth, real-time interactive analysis experience.

[0045] First, this step proactively and selectively schedules data based on the user's analytical intent.

[0046] In a specific application scenario, taking neuroscience research as an example, during the encoding of specific memories in neuroscience research, the neural activity pathway from the CA1 region of the hippocampus to the prefrontal cortex is analyzed. The scientist inputs the command via voice or gesture controller: "Trace the excitatory pathway from the CA1 region of the hippocampus to the prefrontal cortex." The rendering engine deployed in the cloud has a built-in command parsing module (e.g., a Natural Language Processing (NLP) module). This module parses the command, identifying the two key entities, "CA1 region of the hippocampus" and "prefrontal cortex" (i.e., the macroscopic partitions in step S101), and the connectivity attribute "excitability." The rendering engine then generates a high-priority data request. The command data streaming module loads the high-LOD level data (LOD 2) related to these two regions with the highest priority and highlights the relevant data connections based on the connectivity attributes. This rendering method no longer passively loads data based on the user's visual cone, but instead schedules data according to the user's analytical intent, avoiding wasting rendering resources on massive amounts of irrelevant data.

[0047] Furthermore, in order to achieve a near-zero latency interactive experience, this method also includes a predictive data loading mechanism based on a domain model.

[0048] Please see Figure 3This diagram illustrates the principle of predictive loading. Taking neural connectome data as an example, when a user activates a functional cluster (current interaction area A) in VR using a virtual probe, the predictive loading mechanism is activated. This mechanism utilizes a pre-defined domain model that describes the interaction patterns between data units; in this case, it is a computationally lightweight Leaky integrate-and-fire (LIF) neuron model, whose membrane potential V(t) changes with time t according to the differential equation: τ_m * dV(t) / dt = -(V(t) - V_rest) + R_m * I(t), where τ_m is the membrane time constant, V_rest is the resting potential, R_m is the membrane resistance, and I(t) is the input current. The system quantifies the activity of the currently activated cluster A, for example, as the average firing frequency of neurons within the cluster, and converts it into the input current I(t) for all downstream connected clusters (B and C in the diagram) through a pre-defined scaling function, using this as the input to the LIF model. Subsequently, the system performs a forward-looking simulation in the background within a very short time window (e.g., 100-200 milliseconds). The simulation results show that the membrane potential of cluster B is expected to reach the firing threshold after 150 milliseconds, while that of cluster C will not. Based on this prediction, the rendering engine immediately and asynchronously loads the high-LOD level data related to cluster B into the GPU cache in the background. In this way, when the user's exploration naturally moves from A to B along the inherent data correlation, the required high-definition data is already ready, achieving a visually seamless and latency-free transition of details. As an alternative, the prediction model here can also adopt models from other domains. For example, in digital city applications, it could be a traffic flow prediction model to preload a high-precision model of the next block the user is most likely to visit.

[0049] Step S3: Apply a priority synchronization strategy for collaborative interaction.

[0050] This step ensures logical consistency of key operations and accuracy of data synchronization when multiple users collaborate in the same VR environment.

[0051] Specifically, the steps include:

[0052] Step S301: Establish a priority sorting mechanism that is semantically bound to the interaction operation type.

[0053] In a specific implementation, taking a collaborative research scenario as an example, the system predefines a priority list for operation commands. For instance: analytical operations with the highest priority include triggering global simulations (such as signal propagation simulations) and modifying shared global parameters (such as modifying shared neural circuit parameters or synaptic weight parameters); annotation operations with medium priority include adding or deleting a 3D annotation visible to all users and measuring and sharing the size of a data unit; and navigation operations with lower priority include moving the user's personal viewpoint and updating the position and pose of the avatar. When the collaborative synchronization server receives operation commands from different users, it immediately assigns priorities to the commands according to this list.

[0054] Step S302: Based on the priority of the operation, a differential synchronization mechanism is used to synchronize data with the specified network channel.

[0055] Please see Figure 4 The diagram illustrates the differential synchronization mechanism. The system establishes at least two network communication channels. Channel 1 is a high-reliability channel, such as based on TCP or a custom UDP that implements reliable transmission, ensuring that data packets arrive in an ordered and lossless manner. Channel 2 is a standard, best-effort, unreliable channel, such as standard UDP, allowing for a certain degree of packet loss and out-of-order delivery. When the collaborative synchronization server processes operation instructions: for high-priority analytical operations, the instructions and result data must be broadcast through the high-reliability channel 1 to ensure that all clients receive and execute the critical operation in the same order, thereby guaranteeing the consistency of the global simulation results. For medium-priority annotation operations, they can also be transmitted through channel 1. For low-priority navigation operations, the high-frequency position update data is broadcast through the unreliable channel 2. Even if a few data packets are lost in the middle, it will only cause a slight fluctuation in the virtual avatar seen by other users, without affecting the core analytical data and logic. This method, by treating the synchronization requirements of different operations differently, ensures the rigor of the core analysis while avoiding bandwidth waste caused by synchronizing all states, achieving efficient and reliable professional-level collaborative work.

[0056] Example 2

[0057] This embodiment provides a VR large-space rendering system based on a cloud server cluster, which is a hardware and functional modular implementation of the method described in Embodiment 1. Please refer to... Figure 5 The system includes:

[0058] The data preprocessing module 101 is configured to perform the data partitioning and multi-scale level-of-detail construction based on domain prior knowledge as described in step S1 of embodiment 1. In physical implementation, this module can be an offline computing cluster, such as using a big data processing framework like Apache Spark or Hadoop, running customized graph partitioning and LOD generation algorithms to process the raw data and store it in a distributed database or object storage.

[0059] The cloud rendering engine module 102, deployed on a cloud server cluster, is configured to perform the analysis intent-driven rendering and predictive data loading described in step S2 of Embodiment 1. This module is the core of the system and typically runs on server nodes equipped with high-performance graphics processing units (GPUs). It integrates an instruction parsing submodule for resolving user commands, a domain model prediction submodule for performing predictive loading, and a graphics rendering pipeline responsible for real-time 3D rendering and video stream encoding.

[0060] The collaborative synchronization server module 103 is configured to execute the application-oriented priority synchronization strategy for collaborative interaction described in step S3 of embodiment 1. This module is an independent network service responsible for managing all connected VR client sessions, maintaining the shared virtual scene state, receiving and distributing user operation commands, and synchronizing data through different network channels according to preset priority rules.

[0061] And one or more virtual reality clients 104, which can be applications running on mainstream VR headsets (such as HTC Vive, Oculus Quest, etc.). These clients are configured to present users with high-definition video streams encoded and streamed by the cloud rendering engine module, capture user head and controller movements, voice commands, and other interactive inputs, and send these inputs to the cloud rendering engine module and the collaborative synchronization server module, while simultaneously receiving and applying shared state updates from the collaborative synchronization server.

[0062] Those skilled in the art will understand that the interactions between the modules of the aforementioned system are collaborative. For example, the VR client sends the user's voice command "tracking path" to the cloud rendering engine. After parsing, the engine requests high LOD data from the data preprocessing module (or its generated database), renders the data, and then sends the video stream back to the client. Simultaneously, if this operation needs to be visible to other collaborators, relevant status information is sent to the collaboration synchronization server, which then broadcasts it to other clients.

[0063] Example 3

[0064] This embodiment provides another specific application of the method of the present invention in the field of digital twin cities. A digital twin city model is a typical large-scale and complex three-dimensional data, containing detailed geospatial information such as buildings, roads, pipelines, and vegetation.

[0065] First, the data organization method based on domain prior knowledge described in step S1 is applied to the digital twin model data of the city. In this scenario, the domain prior knowledge specifically refers to the city's geospatial planning information.

[0066] When performing macro-level pre-zoning, the city's administrative divisions (e.g., districts, streets) and functional area planning (e.g., central business district, residential area, industrial area) are used as domain-specific spatial zoning standards to initially divide the model data of the entire city into different macro-level data blocks.

[0067] When performing functional cluster identification, a domain-weighted clustering algorithm is used within each macro-level data block (e.g., a central business district). The weight w(i,j) comprehensively considers the functional association between two buildings i and j (e.g., belonging to the same large shopping mall), building type similarity, and spatial proximity. This algorithm can identify functionally closely related building groups (e.g., a complete set of subway station entrances, attached shops, and plazas) as a functional cluster.

[0068] The LOD model is constructed based on the partitioning and clustering results. At the lowest level of detail (LOD 0), only abstract color blocks representing each functional area are rendered; at the medium level of detail (LOD 1), simplified outline models of functional clusters and main roads are rendered; at the highest level of detail (LOD 2), high-precision models of individual buildings, street facilities, and vegetation with detailed textures are rendered.

[0069] Secondly, when interacting with the digital twin city, the analysis intent-driven rendering and predictive loading described in step S2 are performed.

[0070] When executing intent-driven rendering, if city managers input the command "Analyze traffic congestion on Main Road X during the morning rush hour," the system's command parsing module identifies the key entity "Main Road X" and the event "Morning rush hour congestion." The rendering engine will immediately load and render high-LOD level data of "Main Road X" and all buildings and roads within a one-kilometer radius around it with high priority, and may highlight congested road sections based on traffic simulation data.

[0071] When performing predictive data loading, it utilizes a traffic flow prediction model as its domain model. When a user simulates driving a car along road A in a VR environment, the model can predict, based on real-time traffic light status, road topology connectivity, and a pre-defined driving behavior model, that the vehicle is most likely to enter road B at the next intersection. Based on this prediction, the rendering engine asynchronously loads a high-LOD model of the road B region in the background beforehand, thus ensuring a smooth roaming experience when the virtual car actually turns into road B.

[0072] Finally, in scenarios where multiple users collaborate on urban planning reviews, the priority synchronization strategy for collaborative interaction described in step S3 is applied.

[0073] For high-priority analytical operations, such as a planner triggering a global "sunshine analysis" or "urban wind corridor simulation" calculation, or modifying the plot ratio parameters of a shared land, the instructions and results are synchronized through a highly reliable channel to ensure that all participants see consistent simulation results and planning parameters.

[0074] For low- to medium-priority operations, such as a user adding a 3D annotation to "suggest building green space", or updating the virtual avatar positions and viewpoints of various planners in the city model, reliable or unreliable channels can be used for synchronization according to their importance.

[0075] Through the above steps, this invention can efficiently and smoothly support real-time visualization and collaborative analysis of large-scale VR applications such as digital twin cities.

[0076] Example 4

[0077] This embodiment verifies the technical advantages of the method of the present invention through comparative experiments. Please refer to [link / reference]. Figure 7 The figure shows the performance comparison test results of the VR large-space rendering method based on cloud server clusters compared with existing technologies.

[0078] In rendering latency performance tests, such as Figure 7 As shown in sub-figure (a), when processing large-scale complex 3D data ranging from 10GB to 10TB, the rendering latency of the method of the present invention only increases from 15 milliseconds to 70.3 milliseconds, exhibiting a logarithmic growth characteristic. In contrast, the latency of traditional single-machine VR rendering methods surges from 20 milliseconds to 8020 milliseconds under the same data scale, and the latency of general cloud rendering technology increases from 25 milliseconds to 3273 milliseconds. The method of the present invention is 114 times faster than traditional methods and 46 times faster than general cloud rendering under a data scale of 10TB.

[0079] In bandwidth utilization efficiency tests, such as Figure 7As shown in subgraph (b), as the number of concurrent users increases from 1 to 200, the bandwidth utilization efficiency of the method of this invention remains above 0.70, and maintains a high efficiency of 0.83 even with 200 concurrent users. The efficiency of traditional single-machine VR rendering methods decreases linearly from 0.6 to 0.2, while the efficiency of general cloud rendering technology decreases rapidly from 0.7 to near 0.

[0080] In interactive response time testing, such as Figure 7 As shown in subgraph (c), as the scene complexity increases from simple to extremely complex, the response time of the method in this invention only increases from 8 milliseconds to 13.3 milliseconds, the smallest increase. The response time of traditional single-machine VR rendering methods increases from 15 milliseconds to 50 milliseconds, and that of general cloud rendering technology increases from 12 milliseconds to 43 milliseconds.

[0081] In the comprehensive performance index test, such as Figure 7 As shown in (d), the comprehensive performance index of the method of the present invention reaches 95, 92, 88 and 85 points respectively in four application scenarios: small-scale single user, medium-scale small number of users, large-scale medium user, and ultra-large-scale large number of users. It is superior to the traditional single-machine VR rendering method (70, 55, 35 and 20 points) and the general cloud rendering technology (75, 65, 50 and 40 points) in all test indicators.

[0082] The experimental results above demonstrate the technical advantages of the method of the present invention in processing large-scale complex 3D data, especially its performance under challenging conditions such as large data scale, high concurrency users, and complex scenarios.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A VR large-space rendering method based on a cloud server cluster, characterized in that, The method includes: Step one: Based on domain prior knowledge, partition the large-scale complex 3D data and construct multi-scale Level of Detail (LOD). The construction includes: The large-scale complex 3D data is macroscopically pre-partitioned using domain-specific spatial partitioning standards, dividing the data into macroscopic data blocks corresponding to different domain partitions; Within each macro-level data block, functional clusters are identified using community detection or clustering algorithms weighted by domain characteristics; and Based on the partition and cluster identification results, a LOD model including multiple levels of detail is constructed; Step two involves executing intent-driven rendering and predictive data loading, including: when the rendering engine deployed on the cloud server cluster receives a user analysis command from the virtual reality client, parsing the semantics of the command, identifying the key data partitions and functional clusters involved in the command, and streaming and rendering high-LOD level data related to the key data partitions and functional clusters with high priority; and A predictive data loading mechanism based on a domain model is initiated. This mechanism utilizes a pre-defined domain model that can describe the interaction or state transition patterns between data units. Based on the state of the functional clusters currently observed or activated by the user, it performs forward-looking behavior prediction, calculates the downstream data unit clusters most likely to be affected or activated within a preset time window, and asynchronously preloads high LOD level data related to the downstream data unit clusters most likely to be affected or activated into the rendering cache. Specifically, the forward-looking behavior prediction includes: quantifying the state of the currently observed or activated data unit cluster into the average activity intensity of the data units within the cluster, converting it into input stimulus for the downstream connected cluster through a preset scaling function, and using this as the input to the domain model; performing forward-looking simulation calculations through the domain model, and determining the downstream data unit cluster most likely to be affected or activated based on whether the state parameters of the downstream data unit cluster in the simulation results reach a preset activation threshold; Step 3: Apply a priority synchronization strategy for collaborative interaction, including: establishing a priority sorting mechanism that is semantically bound to the type of interaction operation, classifying user operation instructions into analytical operations with the highest priority, commentary operations with the medium priority, and navigation operations with the lower priority. as well as Based on the priority of the operations, a differential synchronization mechanism is used for data synchronization. Specifically, analytical operations are synchronized through a high-reliability data channel, while navigational operations are synchronized through a non-reliable channel.

2. The method according to claim 1, characterized in that, The method of using a domain-specific spatial partitioning standard to perform macroscopic pre-partitioning of the large-scale complex 3D data specifically includes: mapping the spatial coordinates of each data unit to the partition corresponding to the partitioning standard, assigning a partition label to each data unit, thereby dividing the data into multiple macroscopic data blocks corresponding to different domain partitions.

3. The method according to claim 1, characterized in that, The community detection or clustering algorithm based on domain characteristics specifically includes: for the connection between data units i and j, the weight w(i,j) is defined as a weighted sum of multiple domain characteristics, and its calculation formula is: w(i,j)=α*S(i,j)+β*T(i,j)+γ / D(i,j), where S(i,j) is the standardized connection strength, T(i,j) is the data unit type similarity function, D(i,j) is the spatial distance between data units, and α, β, and γ are preset weight coefficients, and the sum of α, β, and γ is 1, thereby constructing a weighted graph with neurons as nodes and the weights as connection edge weights; and running the community detection or clustering algorithm on this weighted graph to identify functionally closely connected data unit clusters.

4. The method according to claim 1, characterized in that, The construction of the LOD model specifically includes: at the lowest level of detail, rendering abstract geometry representing macroscopic partitions; at the medium level of detail, rendering graphics composed of the functional clusters and hyperedges connecting different functional clusters; and at the highest level of detail, resolving the hyperedges into high-resolution individual data connections or data unit shapes.

5. The method according to claim 1, characterized in that, The aforementioned forward-looking behavior prediction specifically includes: using the state of the currently observed or activated data unit cluster as input to the domain model, and using the domain model to calculate and predict the downstream data unit cluster that is most likely to be affected or activated within the preset time window.

6. The method according to claim 5, characterized in that, The large-scale complex three-dimensional data is neural connectome data, the data unit is a neuron, and the domain model is the Leakyintegrate-and-fire model. Taking the state of the currently observed or activated data unit cluster as input specifically includes: simulating the activity of the currently activated neuron cluster as an input current to all downstream connectome clusters, and calculating the change of the neuronal membrane potential of the downstream connectome clusters over time according to the Leakyintegrate-and-fire model to determine the downstream neuron clusters most likely to be activated.

7. The method according to claim 1, characterized in that, The process of parsing the semantics of the instruction specifically includes: using a natural language processing module to parse the user analysis instruction in order to identify one or more key entities contained in the instruction and the connection attributes between entities.

8. The method according to claim 1, characterized in that, The analytical operations include: triggering a global simulation operation or modifying shared global parameters.

9. The method according to claim 1, characterized in that, The high-reliability data channel is a data channel that ensures data packets arrive in an orderly and lossless manner, while the non-reliable channel is a data channel that allows data packets to be lost or out of order.

10. A VR large-space rendering system based on a cloud server cluster, characterized in that, The system includes: The data preprocessing module is configured to perform partitioning and multi-scale Level of Detail (LOD) construction of large-scale complex 3D data based on domain prior knowledge. The construction includes: The large-scale complex 3D data is macroscopically pre-partitioned using domain-specific spatial partitioning standards, dividing the data into macroscopic data blocks corresponding to different domain partitions; Within each macro data block, community detection or clustering algorithms based on domain characteristics are used to identify functional clusters; and based on the partitioning and cluster identification results, a Level of Detail (LOD) model including multiple levels of detail is constructed. The cloud rendering engine module, deployed on a cloud server cluster, is configured to perform analytics intent-driven rendering and predictive data loading, including: When the rendering engine deployed on the cloud server cluster receives a user analysis command from a virtual reality client, it parses the semantics of the command, identifies the key data partitions and functional clusters involved, and then streams and renders high-LOD level data related to those key data partitions and functional clusters with high priority; and A predictive data loading mechanism based on a domain model is initiated. This mechanism utilizes a pre-defined domain model that can describe the interaction or state transition patterns between data units. Based on the state of the data unit cluster currently observed or activated by the user, it performs forward-looking behavior prediction, calculates the downstream data unit clusters most likely to be affected or activated within a preset time window, and asynchronously preloads high LOD level data related to the downstream data unit clusters most likely to be affected or activated into the rendering cache. The collaborative synchronization server module is configured to execute application-oriented priority synchronization strategies for collaborative interactions, including: establishing a priority sorting mechanism semantically bound to the interaction operation type, classifying user operation commands into analytical operations assigned the highest priority, commentary operations assigned the medium priority, and navigation operations assigned the lowest priority; and Based on the priority of the operations, a differential synchronization mechanism is used for data synchronization. Specifically, analytical operations are synchronized through a high-reliability data channel, while navigational operations are synchronized through a non-reliable channel.

Images