Low-altitude real scene three-dimensional database dynamic updating and scenario analysis method and system

Abstract

This invention relates to the field of low-altitude economic technology, and discloses a method and system for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database. The method includes: constructing a basic scene field based on neural implicit representation, mapping 3D spatial coordinates into multi-dimensional feature vectors containing geometry, texture, and semantics; acquiring multi-source real-time perception data, locating changed areas through change detection, and using incremental learning to fine-tune local parameters of the basic scene field to achieve semantic-level dynamic updates; and constructing a hierarchical cognitive architecture comprising a cloud-based "brain" and an edge-based "cerebellum," used for global semantic understanding and macro-path planning, as well as real-time obstacle avoidance and micro-maneuver decision-making, respectively. This invention achieves near real-time updating and deep cognition of the low-altitude database, significantly reducing storage and transmission overhead, improving the safety and collaborative efficiency of high-density low-altitude flights, and providing a self-evolving intelligent digital foundation for the low-altitude economy.

Description

Technical Field

[0001] This invention relates to the field of low-altitude economic technology, specifically to a method and system for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database. Background Technology

[0002] With the rapid development of low-altitude economic applications such as drone logistics, urban air traffic, and emergency rescue, unprecedented demands have been placed on the digital representation of low-altitude resources. As the digital foundation for low-altitude flight, the real-scene 3D database directly determines the safety and economy of low-altitude operations through its data freshness, update efficiency, and scene recognition capabilities.

[0003] Traditional real-world 3D modeling primarily relies on methods such as oblique photogrammetry, LiDAR scanning, and manual modeling to construct discretized models based on triangular meshes (TINs) or point clouds. While these methods have achieved good results in static scene reconstruction, they have inherent limitations when dealing with dynamic changes in low-altitude environments: First, the update cycle is long, often taking weeks or even months from data acquisition and processing to storage, failing to meet the real-time requirements of "day-level" or even "hour-level" changes in urban environments such as building construction and temporary obstacle deployment; second, storage and transmission costs are high, with TB-level discrete models difficult to load and query in real time at the edge of the drone; third, the model only records geometry and texture, lacking an active understanding of scene semantics (such as no-fly zones, take-off and landing points, and risk areas), failing to provide cognitive support for airspace management and flight decisions.

[0004] In terms of scenario-based analysis, existing technologies mostly adopt a post-flight data feedback processing model or rely on simple static geofencing for conflict early warning. Some studies have attempted to introduce machine learning for flight path planning, but these typically use a single model or fixed algorithm, lacking a hierarchical cognitive architecture: global planning and local real-time obstacle avoidance are disconnected, making it difficult for macro-level paths to adapt to micro-level emergencies, while micro-level decisions lack guidance from the overall situation. In addition, the current database update and application analysis are a one-way process, and the analysis results cannot actively drive data collection and updates, creating a vicious cycle of "the older the data, the less useful it is; the less useful it is, the less it is updated."

[0005] In recent years, neural implicit representations (such as NeRF and SDF) have shown great potential in the field of 3D reconstruction. They store continuous scene fields with neural network weights, support arbitrary precision queries and differentiable rendering, and provide new ideas for dynamic updates. However, how to apply this technology to the real-time updating of large-scale low-altitude resources and deeply integrate it with scene-based analysis to form an intelligent database system with cognitive and evolutionary capabilities remains a technical challenge that urgently needs to be solved in this field.

[0006] In summary, existing technologies lack a comprehensive solution that enables high-frequency dynamic updates of low-altitude resource scene 3D databases, in-depth scene cognition, and predictive closed-loop analysis. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database, so as to solve the technical problems of long database update cycles and difficulty in meeting the high-frequency dynamic changes of the low-altitude environment in the prior art.

[0008] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0009] A method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database includes the following steps:

[0010] Step 1: Construct a basic scene field based on neural implicit representation, wherein the scene field is generated by a neural network. Parameterization, converting three-dimensional spatial coordinates Mapped to a multidimensional feature vector containing geometry, texture, and semantics. ;

[0011] Step 2: Acquire multi-source real-time sensing data and locate the changed areas using change detection algorithms. Furthermore, an incremental learning mechanism is used to fine-tune the local parameters of the basic scene field, thereby achieving semantic-level dynamic updates.

[0012] Step 3: Construct a hierarchical cognitive architecture, including a multimodal large model deployed in a cloud computing center as the cloud brain, and a lightweight hybrid model deployed on the edge computing unit of the drone as the edge cerebellum. The cloud brain is used for global semantic understanding and macro-path planning of the updated scene, and the edge cerebellum is used for real-time obstacle avoidance and micro-maneuvering decision-making on the aircraft.

[0013] Step 4: Based on the analysis results of Step 3, perform situational simulation of the low-altitude scenario. When the prediction results meet the preset conditions, trigger the data acquisition task and feed the acquired data back to Step 2 to form a predictive update closed loop.

[0014] As a preferred embodiment of the present invention, the neural implicit scene field constructed in step 1 includes: defining a function. ,in This is a position encoding function used to map low-dimensional coordinates to a high-dimensional feature space; The symbolic distance value is used to characterize geometry. Color values ​​are used to characterize texture. A semantic label vector;

[0015] The scene field supports direct and inverse queries with arbitrary precision.

[0016] As a preferred embodiment of the present invention, the semantic-level dynamic update in step 2 specifically includes:

[0017] Step 2.1: Utilize a lightweight differential network Compare with newly acquired two-dimensional images Image from the corresponding viewpoint rendered from the current implicit scene field Generate a probability map of change And based on this, locate the three-dimensional change area. ;

[0018] Step 2.2: Freeze the network parameters corresponding to the unchanged regions in the implicit scene, and only use the new data to modify the changed regions. The network parameters within the corresponding local spatial range are backpropagated and fine-tuned using gradient descent, with the update rule being: ,in For parameter mask, This is the loss function for the new data.

[0019] As a preferred embodiment of the present invention, the multimodal large model input deployed in the cloud computing center as the cloud brain in step 3 includes a global semantic feature map extracted from the implicit field. Task description vector Based on environmental data, output a set of macroscopic path candidate sets. ;

[0020] Its policy network Through imitation learning or reinforcement learning training, macroscopic air corridors that meet task constraints are generated.

[0021] As a preferred embodiment of the present invention, the lightweight hybrid model deployed on the edge computing unit of the UAV in step 3 as the edge endocerebellum includes a graph neural network module and a reinforcement learning module.

[0022] The graph neural network module is used to construct dynamic interactive graphs. And extract its own node features The reinforcement learning module is based on features With local environmental characteristics Output action distribution It achieves millisecond-level obstacle avoidance decision-making.

[0023] As a preferred embodiment of the present invention, the predictive update feedback closed loop in step 4 specifically includes:

[0024] Step 4.1: Based on the cloud-based brain, construct a digital twin simulation environment, conduct large-scale simulation and extrapolation of specific low-altitude scenarios, and generate demand forecast maps for future periods. ;

[0025] Step 4.2: When At that time, it automatically generates a data collection task containing time and spatial coordinates, and dispatches at least one drone to the designated area to perform targeted data collection;

[0026] Step 4.3: Use the collected data as input to perform the semantic-level dynamic update in Step 2.

[0027] As a preferred embodiment of the present invention, the present invention provides a system for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database, applied to a method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database. The system includes:

[0028] Implicit Field Building Module: Used to build and store neural network-based... Differentiable continuous three-dimensional scene field;

[0029] Semantic update engine module: Connects to multi-source data streams to acquire real-time multi-source sensing data and locates changed areas through change detection algorithms. Furthermore, an incremental learning mechanism is used to fine-tune the local parameters of the basic scene field, thereby achieving semantic-level dynamic updates.

[0030] The hierarchical cognitive module includes a cloud-based cognitive submodule and an edge-based cognitive submodule, which are used to construct a hierarchical cognitive architecture. It includes a multimodal large model deployed in the cloud computing center as the cloud brain, and a lightweight hybrid model deployed on the edge computing unit of the UAV as the edge cerebellum. The cloud brain is used to perform global semantic understanding and macro-path planning on the updated scene, and the edge cerebellum is used to perform real-time obstacle avoidance and micro-maneuvering decisions on the aircraft.

[0031] The simulation and feedback control module is used to perform situational simulation of low-altitude scenarios based on the analysis results of the hierarchical cognition module. When the prediction results meet the preset conditions, it triggers the data acquisition task and feeds the acquired data back to the semantic update engine module to form a predictive update closed loop.

[0032] As a preferred embodiment of the present invention, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database.

[0033] As a preferred embodiment of the present invention, the present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] This invention employs neural implicit representation as the underlying encoding method for the database, storing 3D scenes as neural network weights. Compared to traditional discrete triangulation models, this reduces storage volume by 2-3 orders of magnitude, significantly lowering bandwidth pressure for data transmission and edge loading. More importantly, the semantic-level incremental update mechanism proposed in this invention, through change detection and local parameter fine-tuning, only requires network weight updates within minutes for changed areas. This enables accurate modeling of dynamic elements such as newly added obstacles and building changes, shortening the update cycle of traditional methods from weeks or even months to hours, truly achieving near real-time synchronization between low-altitude data and the physical world.

[0036] The "brain-cerebellum collaboration" hierarchical cognitive architecture constructed in this invention cleverly balances the contradiction between global optimization and local real-time performance. The cloud-based "brain" performs global semantic understanding of implicit scene fields based on a multimodal large model, enabling it to plan air corridors that balance safety, efficiency, and energy consumption at a macro level. The edge-based "cerebellum," through graph neural networks and reinforcement learning, completes the perception and avoidance of sudden dynamic obstacles within milliseconds. The collaborative work of the two avoids the computational explosion of a single model under complex tasks while ensuring that micro-level decisions do not deviate from the macro-strategy, significantly improving the safety and collaborative efficiency of high-density, high-complexity low-altitude flight.

[0037] This invention breaks through the traditional unidirectional "collection-processing-application" process and innovatively proposes a predictive update mechanism. Through simulation and extrapolation of future low-altitude situations, the system can proactively identify data gaps and potential risks, and automatically allocate resources for targeted collection, transforming database updates from "passive response" to "proactive prediction." This closed-loop mechanism ensures that limited collection resources are invested in the most valuable areas, allowing the database to continuously optimize and evolve through ongoing application, providing a qualitative leap from "static maps" to "live aerial maps" for scenarios such as low-altitude traffic management, emergency rescue, and logistics distribution. Attached Figure Description

[0038] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0039] Figure 1 This is a flowchart of the method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database provided in an embodiment of the present invention;

[0040] Figure 2 This is a block diagram of a low-altitude real-scene 3D database dynamic update and scene-based analysis system provided in an embodiment of the present invention;

[0041] Figure 3 A performance result comparison chart provided for embodiments of the present invention. Detailed Implementation

[0042] 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 embodiments of the present invention, and not all embodiments. Based on the 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.

[0043] like Figure 1 As shown, this invention provides a method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database, comprising the following steps:

[0044] Step 1: Construct a basic scene field based on neural implicit representation. The scene field is constructed by a neural network. Parameterization, converting three-dimensional spatial coordinates Mapped to a multidimensional feature vector containing geometry, texture, and semantics. ;

[0045] Step 2: Acquire multi-source real-time sensing data and locate the changed areas using change detection algorithms. Furthermore, it utilizes an incremental learning mechanism to fine-tune local parameters of the basic scene field, thereby achieving semantic-level dynamic updates.

[0046] Step 3: Construct a hierarchical cognitive architecture, including a multimodal large model deployed in a cloud computing center as the cloud brain, and a lightweight hybrid model deployed on the edge computing unit of the drone as the edge cerebellum. The cloud brain is used for global semantic understanding and macro-path planning of the updated scene, while the edge cerebellum is used for real-time obstacle avoidance and micro-maneuvering decisions on the aircraft.

[0047] Step 4: Based on the analysis results of Step 3, perform situational simulation of the low-altitude scenario. When the prediction results meet the preset conditions, trigger the data acquisition task and feed the acquired data back to Step 2 to form a predictive update closed loop.

[0048] The neural implicit scene constructed in step 1 includes: defining functions ,in This is a position encoding function used to map low-dimensional coordinates to a high-dimensional feature space; The symbolic distance value is used to characterize geometry. Color values ​​are used to characterize texture. A semantic label vector;

[0049] Among them, the scene field supports direct query and inverse query with arbitrary precision.

[0050] This embodiment specifically describes one implementation method of steps S1 and S2, as follows:

[0051] First, this invention constructs an implicit scene field that jointly encodes geometric, textural, and semantic information, the core of which is a multilayer perceptron (MLP) network. .

[0052] To improve the ability to represent high-frequency details, the input three-dimensional coordinates are first positionally encoded:

[0053] ;

[0054] in In three-dimensional space coordinates, The number of layers for encoding frequency (in this embodiment, we take...) Positional encoding maps low-dimensional inputs to a high-dimensional space, making it easier for the network to learn high-frequency geometry and textures.

[0055] network With the encoded coordinates As input, output a multidimensional feature vector: ,in:

[0056] Point The signed distance to the nearest object surface: positive values ​​represent the outside, negative values ​​represent the inside, and zero represents the surface. This represents the RGB color value of that point; This indicates that the point belongs to The probability distribution of semantic categories (such as buildings, vegetation, obstacles, no-fly zones, etc.).

[0057] Differentiable rendering technology can render depth maps, color maps, and semantic segmentation maps from any viewpoint, thereby enabling comparison with real images.

[0058] In one specific implementation, The system employs a multilayer perceptron (MLP) architecture, specifically comprising eight fully connected layers (hidden layers), each with 256 neurons. The first seven hidden layers are followed by a ReLU activation function. The final layer splits into three branches based on the output features: a geometric branch outputting the signed distance. (No activation function), texture branch outputs RGB colors (Sigmoid activation function), semantic branch outputs semantic probability (Softmax activation function). This network structure design enables the same set of network parameters to simultaneously encode geometric, textural, and semantic information, achieving alignment and association of the three in implicit space.

[0059] In the initial construction phase, this invention utilizes existing 3D models or multi-view images to pre-train the network. Input data (3D point coordinates) and output supervision signal (true distance) ,color Semantic tags There is a direct physical correspondence: when the point coordinates are located on the surface of the object, , The true color of that point. This represents the true semantic category of the point. This inherent correlation between input and output ensures that the model learns a physically accurate representation of the 3D scene.

[0060] The loss function is defined as:

[0061] ;

[0062] in:

[0063] It is a geometric reconstruction loss, which typically employs the zero level set constraint of the signed distance function at surface points: ,in These are sampling points on a known surface.

[0064] This is the color rendering loss, calculated by comparing the mean squared error between the rendered image and the real image: , To render pixel colors, The actual color.

[0065] It uses semantic segmentation loss, employing cross-entropy: .

[0066] It is the Eikonal regularization term, used to ensure the smoothness of the signed distance field: .

[0067] To balance the weighting coefficients of each loss, in this embodiment, the values ​​are 1.0, 1.0, 0.5, and 0.1, respectively.

[0068] The semantic-level dynamic update in step 2 specifically includes:

[0069] Step 2.1: Utilize a lightweight differential network Compare with newly acquired two-dimensional images Image from the corresponding viewpoint rendered from the current implicit scene field Generate a probability map of change And based on this, locate the three-dimensional change area. ;

[0070] Step 2.2: Freeze the network parameters corresponding to the unchanged regions in the implicit scene, and only use the new data for the changed regions. The network parameters within the corresponding local spatial range are backpropagated and fine-tuned using gradient descent, with the update rule being: ,in For parameter mask, This is the loss function for the new data.

[0071] In this embodiment, when a newly acquired image (with camera pose) When it arrives, first from the current implicit field Render images from the same viewpoint semantic graph Then, a lightweight difference network is used. Calculate the differential heatmap: ;

[0072] in For the parameters of the difference network, This represents the probability that each pixel belongs to a changing region. The differential network uses a U-Net structure, taking the stitched image of two images as input and outputting pixel-level change probabilities.

[0073] To locate changing regions in 3D space, the inverse process of differentiable rendering is used to back-project the pixel-level change probabilities into 3D space, obtaining the spatial bounding box of the changing regions. Specifically, high-probability pixels are converted into 3D point clouds using depth information, and their axis-aligned bounding boxes are calculated.

[0074] This embodiment determines the area of ​​change. Next, we perform fine-tuning of local parameters. We define a binary mask function. ,when If the condition is met, use 1; otherwise, use 0. During fine-tuning, we freeze the network. Most parameters are only updated with The relevant neuron weights. This can be achieved by applying a gradient mask during backpropagation:

[0075] ;

[0076] in It is the learning rate (0.001 in this example). It is a parameter A binary mask of the same dimension indicates which parameters need to be updated (determined by spatial location, through gradient positioning using differentiable rendering). This represents element-wise multiplication. It is a loss function only for newly acquired data (similar in form to the pre-training loss, but only calculated on the newly acquired data). Region-related pixels / points).

[0077] In this way, accurate modeling of new obstacles (such as tower cranes) can be completed in minutes without retraining the entire scene.

[0078] The multimodal large model deployed in the cloud computing center as the cloud brain in step 3 includes the global semantic feature map extracted from the implicit field. Task description vector Based on environmental data, output a set of macroscopic path candidate sets. ;

[0079] Its policy network Through imitation learning or reinforcement learning training, macroscopic air corridors that meet task constraints are generated.

[0080] In this embodiment, the global cognitive model of the cloud-based "brain" adopts a multimodal large model based on Transformer, and its input features are tightly coupled with the requirements of low-altitude management technology.

[0081] Global semantic feature map : By using a regular grid in three-dimensional space ( Dense query implicit field We obtain the feature vector of each grid point. It not only includes geographical and geometric information, but also integrates higher-order semantic information such as no-fly zones and temporary take-off and landing points, providing a cognitive foundation for subsequent macro-planning.

[0082] Task description vector The starting point of the task is encoded. ),end( ), priority ( ) and time window ( ).

[0083] Environmental data: including real-time meteorological data (Wind speed, wind direction, visibility) and dynamic traffic flow data .

[0084] The brain model uses self-attention mechanisms to process the aforementioned multimodal inputs, uncovering deep association rules such as "a certain airspace needs to be detoured due to weather conditions at a certain time" or "an altitude layer needs to be adjusted due to dense traffic flow," ultimately outputting a macroscopic path candidate set. This approach, which inputs implicit field semantics, task constraints, and dynamic environment data into the Transformer for joint reasoning, demonstrates the mutual support between algorithmic features (attention mechanism) and technical features (spatial situation), jointly solving the technical problem of globally optimal path planning.

[0085] The brain model outputs a set of macroscopic path candidates. Each path It is a sequence of three-dimensional waypoints. The model can be trained through imitation learning (learning from expert trajectories) or reinforcement learning (rewarded by task completion efficiency and safety).

[0086] Formally, path generation can be modeled as a sequence decision problem:

[0087] ;

[0088] in It is a policy network. For the first The first path 1 waypoint.

[0089] The lightweight hybrid model deployed on the edge computing unit of the UAV in step 3 as the edge endocerebral system includes a graph neural network module and a reinforcement learning module.

[0090] The graph neural network module is used to build dynamic interactive graphs. And extract its own node features The reinforcement learning module is based on features With local environmental characteristics Output action distribution It achieves millisecond-level obstacle avoidance decision-making.

[0091] In this embodiment, the real-time decision-making model of the edge cerebellum is deployed on the edge computing unit of the UAV, receiving local implicit field compressed packets (i.e., a subset of network parameters related to the current flight area) from the brain. This includes real-time sensor data (such as airborne vision, radar, and ADS-B information). The core of the cerebellum is a hybrid model combining a graph neural network (GNN) and reinforcement learning. At time t... Build dynamic interactive graphs :

[0092] Node features include not only the positions of the machine and surrounding objects obtained through onboard sensors ,speed It also includes information from local implicit fields. Static environmental hazard potential energy obtained through real-time query (For example, if a query point is located in a no-fly zone, its potential energy is infinite; if it is near a building, its potential energy is higher.) This design integrates prior static map knowledge with real-time dynamic perception into the same graph structure.

[0093] side The edge features represent the relative relationships between nodes, including relative distance. Relative velocity Etc. GNN updates node features through message passing:

[0094] ;

[0095] go through After layering, the aggregation features of its own nodes are obtained. Then, the reinforcement learning policy network Output motion distribution (e.g., desired acceleration and steering angle): ,in The static environmental features (such as whether there are obstacles ahead, distance to no-fly zone boundaries, etc.) are queried from the local implicit field. The policy network is trained using the Proximal Policy Optimization (PPO) algorithm, and the reward function is... It takes into account factors such as maintaining a safe distance, mission progress, and energy consumption.

[0096] GNN updates node features through message passing, and after three GraphSAGE layers, it finally outputs the node's own features. It aggregates the motion trends of surrounding dynamic obstacles with the constraints of the static environment. Then, a reinforcement learning policy network... Based on features With local environmental characteristics (If there is a signal blind zone 20 meters ahead, as determined by implicit field query) Output action distribution (Desired acceleration and steering angle). Here and Together, they constitute a complete description of the "environmental state." The obstacle avoidance decisions made by the policy network based on this state are the result of a deep integration of algorithms (GNN, reinforcement learning) and technologies (dynamic obstacle perception, static map query), which solves the technical problem of achieving millisecond-level safe maneuvering in complex low-altitude environments.

[0097] The predictive update feedback loop in step 4 specifically includes:

[0098] Step 4.1: Construct a digital twin simulation environment based on the cloud-based brain, conduct large-scale simulation and extrapolation of specific low-altitude scenarios, and generate demand forecast maps for future periods. ;

[0099] Step 4.2: When At that time, it automatically generates a data collection task containing time and spatial coordinates, and dispatches at least one drone to the designated area to perform targeted data collection;

[0100] Step 4.3: Use the collected data as input to perform the semantic-level dynamic update in Step 2.

[0101] After completing daily simulations, the cloud-based "brain" generates a future time period. Internal demand forecasting chart Each grid value represents the urgency of the area needing an update (e.g., probability of a signal blind spot, likelihood of a new obstacle). Define the update trigger conditions:

[0102] ;

[0103] in The preset threshold is denoted by I, which is an indicator function; in this embodiment, it is set to 0.7. When the triggering condition is met, the system automatically generates a data acquisition task: the target area is... The corresponding three-dimensional spatial range is determined, and idle drones are dispatched to collect the data.

[0104] After the data collection task is completed, the new data will be integrated into the implicit field according to the above incremental update method to complete the closed loop.

[0105] This embodiment uses a 5km×5km low-altitude environment in the central area of ​​a city as the test object to verify the effectiveness of the method of the present invention.

[0106] Dataset: Low-confidence data comes from weekly oblique photogrammetry data collected by routine inspection drones (approximately 500GB / time), while high-confidence data comes from lidar scan data of key areas (approximately 50GB / time). The training set contains 10 historical data collections, and the test set consists of data from 5 newly added temporary obstacles (tower cranes, balloons, etc.) in week 11.

[0107] Setting parameters for implicit field networks An 8-layer MLP with 256 neurons per layer is used, with input position encoding L=10 and a differential network. The system employs a lightweight U-Net with approximately 1 million parameters, a cloud-based brain based on a 12-layer Transformer with 120 million parameters, a limbic cerebellum GNN with 3 layers of GraphSAGE (64 dimensions per layer), and a policy network with 2 layers of MLP.

[0108] To demonstrate the performance advantages of this application, a comparison was made with traditional oblique photogrammetry modeling methods (full reconstruction, one month in cycle), and the results are as follows: Figure 3 As shown, the evaluation metrics include: 1) Update latency: the time from when a change occurs to when the database reflects the change; 2) Storage overhead: the space occupied by the database; 3) Obstacle avoidance success rate: the success rate of the UAV in autonomously avoiding obstacles in an environment with newly added obstacles; and 4) Data utilization rate: the proportion of data used for updates to the total amount of data collected.

[0109] The method of this invention significantly outperforms traditional methods in terms of update latency, storage overhead, and obstacle avoidance success rate. In particular, it requires only 12% of the data to achieve accurate updates. This effect cannot be achieved through simple data aggregation or conventional neural network applications, but rather through the synergistic effect of the semantic-level incremental update mechanism (algorithm feature) and implicit field representation (technical feature) in this invention: the continuity of the implicit field allows for fine-tuning of local parameters, while the semantic-level localization guided by the differential network ensures the accuracy of the fine-tuning. These two mechanisms support each other functionally, jointly bringing about the unexpected technical effect of "achieving near real-time updates with extremely low data utilization." Simultaneously, the "bodies" collaborative architecture improves the obstacle avoidance success rate to 94% when handling high-density conflict scenarios. Compared to the comparison methods, this is not an effect achievable through single model optimization, but rather an overall technical effect resulting from the combined action of global planning and local real-time decision-making algorithms within a hierarchical architecture, working together to control the physical aircraft.

[0110] In summary, this invention systematically addresses the core pain points of existing low-altitude resource databases—namely, delayed updates, cognitive gaps, and application disconnect—by organically combining neural implicit representation, hierarchical cognitive architecture, and predictive update loops. This provides a solid technical foundation for the large-scale, high-safety operation of the low-altitude economy.

[0111] like Figure 2 As shown, this invention provides a system for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database, applied to a method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database. The system includes:

[0112] Implicit field building blocks are used to build and store neural network-based structures. Differentiable continuous three-dimensional scene field;

[0113] The semantic update engine module connects to multi-source data streams to acquire real-time multi-source sensing data and locates changed areas using change detection algorithms. Furthermore, it utilizes an incremental learning mechanism to fine-tune local parameters of the basic scene field, thereby achieving semantic-level dynamic updates.

[0114] The hierarchical cognitive module includes a cloud-based cognitive submodule and an edge-based cognitive submodule, which are used to build a hierarchical cognitive architecture. It includes a multimodal large model deployed in the cloud computing center as the cloud brain, and a lightweight hybrid model deployed on the edge computing unit of the drone as the edge cerebellum. The cloud brain is used to perform global semantic understanding and macro-path planning on the updated scene, while the edge cerebellum is used to perform real-time obstacle avoidance and micro-maneuvering decisions on the aircraft.

[0115] The simulation and feedback control module is used to perform situational simulation of low-altitude scenarios based on the analysis results of the hierarchical cognition module. When the prediction results meet the preset conditions, the data acquisition task is triggered, and the acquired data is fed back to the semantic update engine module to form a predictive update closed loop.

[0116] This invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database.

[0117] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database.

[0118] This invention employs neural implicit representation as the underlying encoding method for the database, storing 3D scenes as neural network weights. Compared to traditional discrete triangulation models, this reduces storage volume by 2-3 orders of magnitude, significantly lowering bandwidth pressure for data transmission and edge loading. More importantly, the semantic-level incremental update mechanism proposed in this invention, through change detection and local parameter fine-tuning, only requires network weight updates within minutes for changed areas. This enables accurate modeling of dynamic elements such as newly added obstacles and building changes, shortening the update cycle of traditional methods from weeks or even months to hours, truly achieving near real-time synchronization between low-altitude data and the physical world.

[0119] The "brain-cerebellum collaboration" hierarchical cognitive architecture constructed in this invention cleverly balances the contradiction between global optimization and local real-time performance. The cloud-based "brain" performs global semantic understanding of implicit scene fields based on a multimodal large model, enabling it to plan air corridors that balance safety, efficiency, and energy consumption at a macro level. The edge-based "cerebellum," through graph neural networks and reinforcement learning, completes the perception and avoidance of sudden dynamic obstacles within milliseconds. The collaborative work of the two avoids the computational explosion of a single model under complex tasks while ensuring that micro-level decisions do not deviate from the macro-strategy, significantly improving the safety and collaborative efficiency of high-density, high-complexity low-altitude flight.

[0120] This invention breaks through the traditional unidirectional "collection-processing-application" process and innovatively proposes a predictive update mechanism. Through simulation and extrapolation of future low-altitude situations, the system can proactively identify data gaps and potential risks, and automatically allocate resources for targeted collection, transforming database updates from "passive response" to "proactive prediction." This closed-loop mechanism ensures that limited collection resources are invested in the most valuable areas, allowing the database to continuously optimize and evolve through ongoing application, providing a qualitative leap from "static maps" to "live aerial maps" for scenarios such as low-altitude traffic management, emergency rescue, and logistics distribution.

[0121] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database, characterized in that, Includes the following steps: Step 1: Construct a basic scene field based on neural implicit representation, wherein the scene field is generated by a neural network. Parameterization, converting three-dimensional spatial coordinates Mapped to a multidimensional feature vector containing geometry, texture, and semantics. ; Step 2: Acquire multi-source real-time sensing data and locate the changed areas using change detection algorithms. Furthermore, an incremental learning mechanism is used to fine-tune the local parameters of the basic scene field, thereby achieving semantic-level dynamic updates. Step 3: Construct a hierarchical cognitive architecture, including a multimodal large model deployed in a cloud computing center as the cloud brain, and a lightweight hybrid model deployed on the edge computing unit of the drone as the edge cerebellum. The cloud brain is used for global semantic understanding and macro-path planning of the updated scene, and the edge cerebellum is used for real-time obstacle avoidance and micro-maneuvering decision-making on the aircraft. Step 4: Based on the analysis results of Step 3, perform situational simulation of the low-altitude scenario. When the prediction results meet the preset conditions, trigger the data acquisition task and feed the acquired data back to Step 2 to form a predictive update closed loop.

2. The method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database according to claim 1, characterized in that, The neural implicit scene constructed in step 1 includes: defining functions. ,in This is a position encoding function used to map low-dimensional coordinates to a high-dimensional feature space; The symbolic distance value is used to characterize geometry. Color values ​​are used to characterize texture. A semantic label vector; The scene field supports direct and inverse queries with arbitrary precision.

3. The method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database according to claim 2, characterized in that, The semantic-level dynamic update in step 2 includes: Step 2.1: Utilize a lightweight differential network Compare with newly acquired two-dimensional images Image from the corresponding viewpoint rendered from the current implicit scene field Generate a probability map of change And based on this, locate the three-dimensional change area. ; Step 2.2: Freeze the network parameters corresponding to the unchanged regions in the implicit scene, and only use the new data to modify the changed regions. The network parameters within the corresponding local spatial range are backpropagated and fine-tuned using gradient descent, with the update rule being: ,in For parameter mask, This is the loss function for the new data.

4. The method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database according to claim 3, characterized in that, In step 3, the input of the multimodal large model deployed in the cloud computing center as the cloud brain includes the global semantic feature map extracted from the implicit field. Task description vector Based on environmental data, output a set of macroscopic path candidate sets. ; Its policy network Through imitation learning or reinforcement learning training, macroscopic air corridors that meet task constraints are generated.

5. The method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database according to claim 4, characterized in that, The lightweight hybrid model deployed on the edge computing unit of the UAV in step 3 as the edge endocerebellum includes a graph neural network module and a reinforcement learning module. The graph neural network module is used to construct dynamic interactive graphs. And extract its own node features The reinforcement learning module is based on features With local environmental characteristics Output action distribution It achieves millisecond-level obstacle avoidance decision-making.

6. The method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database according to claim 5, characterized in that, The predictive update feedback loop in step 4 includes: Step 4.1: Based on the cloud-based brain, construct a digital twin simulation environment, conduct large-scale simulation and extrapolation of specific low-altitude scenarios, and generate demand forecast maps for future periods. ; Step 4.2: When At that time, it automatically generates a data collection task containing time and spatial coordinates, and dispatches at least one drone to the designated area to perform targeted data collection; Step 4.3: Use the collected data as input to perform the semantic-level dynamic update in Step 2.

7. A system for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database, characterized in that, The system, which is applied to the method for dynamic updating and scene-based analysis of a low-altitude real-scene 3D database according to any one of claims 1-6, comprises: Implicit Field Building Module: Used to build and store neural network-based... Differentiable continuous three-dimensional scene field; Semantic update engine module: Connects to multi-source data streams to acquire real-time multi-source sensing data and locates changed areas through change detection algorithms. Furthermore, an incremental learning mechanism is used to fine-tune the local parameters of the basic scene field, thereby achieving semantic-level dynamic updates. The hierarchical cognitive module includes a cloud-based cognitive submodule and an edge-based cognitive submodule, which are used to construct a hierarchical cognitive architecture. It includes a multimodal large model deployed in the cloud computing center as the cloud brain, and a lightweight hybrid model deployed on the edge computing unit of the UAV as the edge cerebellum. The cloud brain is used to perform global semantic understanding and macro-path planning on the updated scene, and the edge cerebellum is used to perform real-time obstacle avoidance and micro-maneuvering decisions on the aircraft. The simulation and feedback control module is used to perform situational simulation of low-altitude scenarios based on the analysis results of the hierarchical cognition module. When the prediction results meet the preset conditions, it triggers the data acquisition task and feeds the acquired data back to the semantic update engine module to form a predictive update closed loop.

Images