Three-dimensional city space data processing method and system for large model

Abstract

The invention discloses a three-dimensional urban spatial data processing method and system for a large model, and mainly solves the problems of poor compatibility between three-dimensional data and an AI model, multi-element splitting, limited static labeling and the like. According to the innovative scheme, the method comprises the following steps: 1, dynamic rasterization: dividing a city space into three-dimensional pixel points according to adjustable granularity, and integrating three-dimensional coordinates and multi-channel attributes; 2, a multi-channel tensor supports multi-channel One-Hot coding of a classification task and three-dimensional bounding box normalization of a target detection task, and an element protection range is automatically expanded based on a rule engine; thirdly, rule-driven label extension is carried out, original values of binary data are reserved, and Min-Max or Z-Score standardization is adopted for continuous data, so that the model training efficiency is improved; and 4, performing channel normalization, including a three-dimensional grid division engine, a multi-source element fusion module, a dynamic label encoder and the like, and realizing full-link connection from data acquisition to AI model training.

Description

Technical Field

[0001] The invention belongs to the field of urban space data processing technology, and in particular to a three-dimensional urban space data processing method and system for large models. Background Art

[0002] With the breakthroughs of AI big models (such as GPT, visual transformer, etc.) in natural language processing and image recognition, their application potential in urban space analysis is gradually emerging. However, the existing technology has significant defects in urban three-dimensional data processing and the adaptability of AI big models, which are as follows: 1. Data dimension fragmentation: The training and output of AI large models usually rely on single-dimensional data (such as text sequences or two-dimensional images), while urban space is essentially a complex system containing multi-dimensional information such as geographic coordinates, height, and facility attributes. Existing technologies lack a method to uniformly encode three-dimensional spatial structure and multi-factor attributes, making it difficult for AI large models to understand the three-dimensional correlation of urban space.

[0003] 2. Inefficient 3D data expression: Traditional methods describe 3D urban space through point cloud, BIM model or oblique photography data, but these data formats are highly redundant, have fixed granularity, and are not directly compatible with the tensor input structure of large AI models. For example, point cloud data lacks semantic attribute channels, and BIM models are difficult to dynamically adjust accuracy, which seriously limits the efficiency of model training.

[0004] 3. Insufficient multi-factor integration capabilities: Urban space includes cross-domain elements such as nature, material, culture, and economy. Existing technologies usually use independent modeling of subsystems (such as GIS systems to process geographic information and BIM systems to process building data), resulting in the fragmentation of multi-source data. AI large models cannot extract cross-factor association rules from heterogeneous data and are difficult to support comprehensive decision-making (such as disaster warnings that require simultaneous analysis of geological, architectural, and pipeline data).

[0005] 4. Poor dynamic adaptability: The boundaries and attributes of urban elements change dynamically over time and in different scenarios (e.g., subway construction requires real-time expansion of the protection range), while existing label encoding methods rely on manual static annotation and cannot automatically adapt to dynamic needs, resulting in limited model generalization capabilities. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a three-dimensional urban space data processing method and system for large models in response to the deficiencies in the above-mentioned prior art. By decomposing the urban space into a three-dimensional grid network with dynamically adjustable particle accuracy, and constructing a multi-dimensional tensor ([C, H, W, L]) that integrates multi-factor attributes, the industry problems of poor compatibility between AI large models and three-dimensional urban data, multi-factor fragmentation and insufficient dynamic adaptability are fundamentally solved.

[0007] The first aspect of the present invention discloses a method for processing three-dimensional urban space data for a large model, comprising the following steps: (a) Divide the target urban space into a three-dimensional grid network according to a preset granularity, where each grid corresponds to a three-dimensional pixel point; (b) determining the types of urban elements intersecting within each three-dimensional pixel point through spatial calculation, wherein the urban elements include natural elements, material elements, cultural elements, and economic elements; (c) Construct a multidimensional urban space pixel tensor with dimensions [C, H, W, L], where: C is the number of channels, each channel corresponds to a city element classification; H, W, and L are the number of three-dimensional pixels in the length, width, and height directions of the urban space, respectively; (d) Label encoding the pixel tensor according to the task type: If it is a classification task, multi-channel one-hot encoding is used to label each 3D pixel point; If it is an object detection task, generate a normalized detection label based on the absolute coordinates of the 3D bounding box; (e) performing channel-by-channel normalization on the pixel tensor, including retaining binary channel data and performing Min-Max normalization or Z-Score normalization on continuous channel data; (f) Generate urban spatial data required for training urban spatial analysis models.

[0008] In the above method, the three-dimensional grid particle size in step (a) is 1m×1m×1m.

[0009] In the above method, the classification task label encoding in step (d) further includes dynamic boundary expansion: Expand the original boundary of the specified city element according to the preset rules to generate the extended area; Mark the three-dimensional pixels in the extended area as a newly added protection category; The expansion rule is: based on the original boundary parameters M(Hm, Wm, Lm), the distance Δ is expanded along the three-dimensional direction to generate the extended boundary M'(Hm±Δ, Wm±Δ, Lm±Δ).

[0010] In the above method, the expansion distance Δ is dynamically set according to the element type, including: Subway protection area: Δ=50m; Municipal pipeline protection area: Δ=20m; Historical building protection area: Δ=30m; The expanded regions are labeled as independent category channels and coexist with the original feature channels in the label tensor.

[0011] In the above method, the target detection task label encoding in step (d) includes: Calculate the absolute coordinates of the 3D bounding box of the target area, including the center point coordinates (x, y, z) and dimensions (dx, dy, dz); Normalize the absolute coordinates to relative coordinates, the formula is: x_center = x / H, y_center = y / W, z_center = z / L; dx_norm = dx / H, dy_norm = dy / W, dz_norm = dz / L; Generates a detection label vector [class_id, x_center, y_center, z_center, dx_norm, dy_norm, dz_norm] where class_id is the class identifier.

[0012] In the above method, the data normalization process in step (e) includes: For binary channel data, the original 0 / 1 value is retained; Min-Max normalization is used for continuous economic factor data, and the formula is: x_norm = (x - x_min) / (x_max - x_min); Save the normalization parameters of the training set and simultaneously apply the same parameters to the test set.

[0013] The second aspect of the present invention discloses an urban space data processing system, comprising: 3D grid partitioning engine, used to divide the urban space into uniform 3D pixels; Multi-source feature fusion module, used to generate pixel tensors containing multi-channel urban features; Dynamic label encoder, which supports one-hot encoding for classification tasks and bounding box normalization for object detection tasks; Channel-wise normalization processor that differentiates and / or normalizes tensor data by channel type.

[0014] In the above system, the dynamic label encoder includes a boundary extension unit, which is used to automatically expand the protection range and update the label tensor according to the element type.

[0015] Compared with the prior art, the present invention has the following advantages: 1. Three-dimensional rasterization and multi-channel fusion: Decompose the urban space into three-dimensional pixels with dynamically adjustable granularity (such as 1m³). Each pixel integrates three-dimensional coordinates (H, W, L) and multi-channel attribute information (such as natural element status, facility function, cultural labels, economic indicators) to form a unified tensor structure [C, H, W, L], which directly adapts to the multi-dimensional input requirements of large AI models.

[0016] 2. Dynamic granularity control: supports flexible adjustment of grid granularity according to task requirements (such as 1m³ for fine facility management and 10m³ for macro planning), achieving a balance between data accuracy and computing efficiency, and overcoming the rigidity of traditional point cloud or BIM models.

[0017] 3. Multi-factor semantic association: By explicitly encoding cross-domain factor attributes through the channel dimension (C), the AI ​​large model can learn the interaction rules between factors (such as the impact of subway vibrations on surrounding buildings) from a unified data source, thereby improving reasoning capabilities in complex scenarios.

[0018] 4. Automatic label extension: Dynamically expand feature boundaries based on the rule engine (for example, the subway protection range automatically extends 50m during the construction phase) and updates pixel labels in real time to solve the model lag problem caused by static labeling and significantly improve the adaptability to dynamic scenes.

[0019] 5. Industrial-grade compatibility: The generated three-dimensional pixel tensor can be directly input into large visual models (such as ViT) and multimodal large models (such as GPT-4V) for end-to-end training, lowering the threshold for AI technology implementation in urban space analysis tasks.

[0020] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The present invention is a flow chart of a method for processing urban space data based on three-dimensional grids.

[0022] Figure 2 This is an example of the initial pixel points in the urban space.

[0023] Figure 3 It is an architecture diagram of the urban space data processing system of the present invention. DETAILED DESCRIPTION

[0024] like Figure 1 As shown, a method for processing three-dimensional urban space data for a large model includes the following steps: Step 1: 3D grid division and data collection (such as Figure 2 shown) Step 1.1: Define the urban spatial scope and granularity Input parameters: Set the three-dimensional size of the target urban space (e.g., length H=1000m, width W=800m, height L=50m) and grid granularity (e.g., 1m×1m×1m); Grid generation: The urban space is divided into 1000×800×50 three-dimensional pixels through the spatial coordinate system, and the coordinates of each pixel are (h,w,l) (h∈[0,1000), w∈[0,800), l∈[0,50)); Dynamic adjustment: If you need to reduce the computational complexity, you can adjust the granularity to 5m×5m×5m and reduce the number of grids to 200×160×10.

[0025] Step 1.2: Collect multi-source urban feature data Data Source: Natural elements: terrain elevation, vegetation distribution, hydrological data (GIS system); Material elements: building BIM model, traffic network, underground pipelines (CAD drawings); Cultural elements: coordinates of historical buildings, protection levels (database); Economic factors: regional GDP, population density (statistical report).

[0026] Spatial calculation: Use spatial indexing technology (such as R-tree) to determine the intersection relationship between each three-dimensional pixel point and urban elements.

[0027] Example: If a pixel point (h,w,l) intersects with a subway tunnel, its material element channel value is marked as 1.

[0028] Step 2: Construct a multidimensional pixel tensor Step 2.1: Channel assignment and data mapping Channel definition: Channel C1 (natural elements): binary label (0 / 1 indicates whether a natural element exists); Channel C2 (material elements): classification code (1 = transportation facilities, 2 = building complex, 3 = municipal pipeline); Channel C3; (Cultural elements): Protection level (0-5); Channel C4 (economic factors): continuous values ​​(such as GDP density) Tensor generation: fill the attributes of each pixel into the corresponding channel to generate a four-dimensional tensor Tensor[C,H,W,L], where C=4,H=1000,W=800,L=50; Step 2.2: Data storage optimization Sparse storage: Sparse matrix compression is used for null value areas (such as underground areas without buildings) to reduce storage overhead.

[0029] Step 3: Label encoding 3.1 Classification Task (Semantic Segmentation) Step 3.1.1: Define the category set Preset categories: natural elements (N), transportation facilities (T), building complex (B), subway protection area (MP), historical building (HB).

[0030] Step 3.1.2: Dynamic Boundary Extension Rules Engine: Input the original boundary parameters (e.g. the center of the subway tunnel Hm=500m, Wm=400m, Lm=10m); The expansion distance Δ=50m generates the expansion area: H′∈[500−50,500+50], W′∈[400−50,400+50], L′∈[10−5,10+5]; Traverse the pixels in the extended area and update the category label from "transportation facilities (T)" to "subway protection area (MP)".

[0031] Step 3.1.3: Generate One-Hot Label Each pixel generates a 5-dimensional vector (corresponding to 5 categories), such as the pixel label of the subway protection range is [0,0,0,1,0].

[0032] 3.2 Object Detection Task (Region Detection) Step 3.2.1: Calculate the 3D bounding box Enter the subway tunnel parameters: center coordinates (500,400,10), dimensions (100m,80m,10m); Normalized calculation: x center =500 / 1000=0.5,y center =400 / 800=0.5, z center =10 / 50=0.2; dx norm =100 / 1000=0.1,dy norm =80 / 800=0.1,dz norm =10 / 50=0.2; Label vector: generates [4, 0.5, 0.5, 0.2, 0.1, 0.1, 0.2], where 4 is the “subway protection area” category ID.

[0033] Step 4: Data normalization Step 4.1: Channel processing Binary channels (such as natural elements): retain the original 0 / 1 value; Continuous channels (such as economic factors): Calculate the global minimum / maximum value of the training set (such as GDP density x min =0,x max =1000); Apply Min-Max normalization: ; Step 4.2: Parameter persistence Save the normalization parameters of the training set (such as x min ,x max ), and apply it synchronously during test set inference.

[0034] Step 5: Model training and validation Step 5.1: Task Adaptation Classification task: Use 3D U-Net model, input tensor [4, 1000, 800, 50], output segmentation result [5, 1000, 800, 50]; Object detection task: Use 3D Faster R-CNN, input the same tensor, and output bounding box coordinates and categories.

[0035] Step 5.2: Dynamic scene verification Test case: Added a new subway construction phase and extended the protection range to Δ=70m; System response: The labels of the extended area were automatically updated, the model was retrained, and the F1-score was improved by 12%.

[0036] Example 1: Dynamic marking of subway protection areas Input: subway tunnel center (500m, 400m, 10m), original size (100m, 80m, 10m), Δ=50m; Output: Classification label: Pixels in the extended area [450−550m, 350−450m, 5−15m] are marked as “MP”; Detection labels: normalized bounding boxes [4, 0.5, 0.5, 0.2, 0.1, 0.1, 0.2].

[0037] Example 2: Multi-factor analysis of historical buildings Input: building center (200m, 300m, 20m), Δ=30m, cultural protection level 3; Output: Classification label: Pixels in the extended area [170−230m, 270−330m, 0−50m] are labeled as “HB”; Economic channel data: Regional GDP density is normalized from 500 to 0.5.

[0038] The technical effect verification is as follows:

[0039] like Figure 3 As shown, the specific implementation method of the urban space data processing system is described below: 1. System Architecture Overview The system consists of the following core modules to support dynamic processing, label encoding and model adaptation of 3D urban spatial data: 3D grid partitioning engine, multi-source element fusion module, dynamic label encoder, channel normalization processor and AI model interface layer; 2. Detailed implementation of the module 2.1 3D grid partitioning engine Function: Cut the urban space into a three-dimensional pixel network with adjustable granularity.

[0040] Hardware implementation: GPU-accelerated parallel computing clusters (such as NVIDIA A100).

[0041] Software Process: 1) Input parameter configuration: Spatial range: absolute coordinates of length (H), width (W), and height (L) (e.g. H=1000m, W=800m, L=50m); Granularity: preset value (such as 1m³) or dynamic rule (such as "1m³ in the city center, 10m³ in the suburbs").

[0042] 2) Grid generation algorithm: Using a spatial hashing algorithm, the three-dimensional coordinates (x,y,z)(x,y,z) are mapped to a unique grid ID:

[0043] Where Δ is the particle size (e.g. 1m).

[0044] 3) Dynamic adjustment mechanism: If it is detected that the density of elements in a certain area exceeds the threshold (such as a densely built-up area), the granularity will be automatically switched from 10m to 1m.

[0045] Example: Input: Urban space 1000m×800m×50m1000m×800m×50m, granularity 1m³; Output: Generate 1000×800×50=40,000,000 three-dimensional pixels, each of which stores the reference coordinates (h, w, l).

[0046] 2.2 Multi-source element fusion module Function: Map multi-source data such as GIS, BIM, statistical reports, etc. to three-dimensional pixel channels.

[0047] Key technologies: spatial relational database (such as PostGIS), semantic parsing engine.

[0048] Processing flow: 1) Data access: Supports multiple input formats: Vector data (Shapefile, GeoJSON); 3D models (IFC, OBJ); Tabular data (CSV, Excel).

[0049] 2) Spatial relationship calculation: Use R-tree index to accelerate spatial intersection judgment; 3) Channel mapping rules: Natural element channel: terrain elevation value is directly filled; Cultural element channel: The protection level of historical buildings is weighted by radius attenuation (e.g., level 5 in the core area and level 3 in the extended area).

[0050] Example: Input: subway BIM model (centerline coordinates + radius); Output: All pixels intersecting the tunnel whose material element channel is marked as "1" (traffic facility).

[0051] 2.3 Dynamic Label Encoder Function: Generate classification or detection labels according to task types, and support rule-driven dynamic expansion.

[0052] Submodules: Classification label unit: One-Hot encoding based on rule engine; Detection label unit: 3D bounding box coordinate normalization; Dynamic expansion engine: automatically adjusts the protection range of elements.

[0053] Implementation details: 1) Rule Engine: Predefined extension rule base (JSON format): "Metro Protection Zone": { "Element Type": "Transportation Facilities", "Extended Distance": {"H": 50, "W": 50, "L": 5}, "New Category": "MP" }, "Historic Building Conservation Area": ​​{ "Element Type": "Cultural Element", "Extended Distance": {"H": 30, "W": 30, "L": 10}, "New Category": "HB" } } Monitor element changes in real time (such as new construction areas) and trigger rule execution.

[0054] 2) Classification label generation: For each pixel, the final category is assigned according to priority (such as MP>HB>building complex) and a One-Hot vector is generated.

[0055] 3) Detection label generation: Call the 3D minimum bounding box algorithm (such as PCA-AABB) to calculate the bounding box and output normalized coordinates.

[0056] Example: Input: original subway area H=500m, W=400m, L=10m, expansion rule Δ=50m; Output: expanded region label [450−550m, 350−450m, 5−15m], detection box [0.5, 0.5, 0.2, 0.1, 0.1, 0.2].

[0057] 2.4 Channel Normalization Processor Function: Process data differently by channel type to improve model training stability.

[0058] Processing logic: 1) Channel type identification: Binary channels (such as feature existence): skip normalization; Continuous channels (such as economic indicators): automatically select Min-Max or Z-Score.

[0059] 2) Distributed computing: Use Apache Spark to parallelize the calculation of channel statistics (such as mean and variance).

[0060] 3) Parameter persistence: Store the normalized parameters of the training set (such as the maximum GDP value of 1000) in the Redis database for inference.

[0061] Example: Input: economic channel original data [0,350,920,1000]; Output: [0, 0.35, 0.92, 1.0] after Min-Max normalization.

[0062] 2.5 AI Model Interface Layer Function: Convert the processed tensor data into AI large model input format.

[0063] Adapter Type: 1) Visual large model adapter: Reshape the tensor [C,H,W,L] into a sequence format supported by ViT (e.g., H×W×L C-dimensional vectors).

[0064] 2) Multimodal large model adapter: Concatenate tensor data with text descriptions (e.g., “subway protected area”) to generate GPT-4V-compatible multimodal inputs.

[0065] Example: Input: 4D tensor [4, 1000, 800, 50]; Output: 1000×800×50=40,000,000 4-dimensional embedding vectors of the ViT input sequence.

[0066] 3. System dynamic collaboration process Scenario: Subway construction leads to changes in the protection scope 1) Data update: The construction party submits new shield section coordinates; 2) Grid redivision: The engine switches the granularity to 1m³ in the changed area; 3) Label extension: The rule engine automatically expands the MP label range; 4) Model hot update: Inject new data into the online model through the interface layer to achieve real-time reasoning.

[0067] 4. Technical effect verification

[0068] Through modular design, rule engine drive and distributed computing, this system realizes full-link automation from the collection of three-dimensional urban spatial data to AI model training, and overcomes key technical bottlenecks such as multi-source heterogeneous data fusion, dynamic scene adaptation and large model compatibility.

[0069] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for processing three-dimensional urban space data for a large model, characterized in that: The following steps are involved: (a) Divide the target urban space into a three-dimensional grid network according to a preset granularity, where each grid corresponds to a three-dimensional pixel point; (b) determining the types of urban elements intersecting within each three-dimensional pixel point through spatial calculation, wherein the urban elements include natural elements, material elements, cultural elements, and economic elements; (c) Construct a multidimensional urban space pixel tensor with dimensions [C, H, W, L], where: C is the number of channels, each channel corresponds to a city element classification; H, W, and L are the number of three-dimensional pixels in the length, width, and height directions of the urban space, respectively; (d) Label encoding the pixel tensor according to the task type: If it is a classification task, multi-channel one-hot encoding is used to label each 3D pixel point; If it is an object detection task, generate a normalized detection label based on the absolute coordinates of the 3D bounding box; (e) performing channel-by-channel normalization on the pixel tensor, including retaining binary channel data and performing Min-Max normalization or Z-Score normalization on continuous channel data; (f) Generate urban spatial data required for training urban spatial analysis models.

2. A method for processing three-dimensional urban space data for a large model according to claim 1, characterized in that: The three-dimensional grid particle size in step (a) is 1m×1m×1m.

3. A method for processing three-dimensional urban space data for a large model according to claim 1, characterized in that: The classification task label encoding in step (d) further includes dynamic boundary expansion: Expand the original boundary of the specified city element according to the preset rules to generate the extended area; Mark the three-dimensional pixels in the extended area as a newly added protection category; The expansion rule is: based on the original boundary parameters M(Hm, Wm, Lm), the distance Δ is expanded along the three-dimensional direction to generate the extended boundary M'(Hm±Δ, Wm±Δ, Lm±Δ).

4. A method for processing three-dimensional urban space data for a large model according to claim 3, characterized in that: The expansion distance Δ is dynamically set according to the element type, including: Subway protection area: Δ=50m; Municipal pipeline protection area: Δ=20m; Historical building protection area: Δ=30m; The expanded regions are labeled as independent category channels and coexist with the original feature channels in the label tensor.

5. A method for processing three-dimensional urban space data for a large model according to claim 1, characterized in that: The object detection task label encoding in step (d) includes: Calculate the absolute coordinates of the 3D bounding box of the target area, including the center point coordinates (x, y, z) and dimensions (dx, dy, dz); Normalize the absolute coordinates to relative coordinates, the formula is: x_center = x / H, y_center = y / W, z_center = z / L; dx_norm = dx / H, dy_norm = dy / W, dz_norm = dz / L; Generates a detection label vector [class_id, x_center, y_center, z_center, dx_norm, dy_norm, dz_norm] where class_id is the class identifier.

6. A method for processing three-dimensional urban space data for a large model according to claim 1, characterized in that: The data normalization process in step (e) includes: For binary channel data, the original 0 / 1 value is retained; Min-Max normalization is used for continuous economic factor data, and the formula is: x_norm = (x - x_min) / (x_max - x_min); Save the normalization parameters of the training set and simultaneously apply the same parameters to the test set.

7. An urban space data processing system, characterized in that: include: 3D grid partitioning engine, used to divide the urban space into uniform 3D pixels; Multi-source feature fusion module, used to generate pixel tensors containing multi-channel urban features; Dynamic label encoder, which supports one-hot encoding for classification tasks and bounding box normalization for object detection tasks; Channel-wise normalization processor that differentiates and / or normalizes tensor data by channel type.

8. The system according to claim 7, characterized in that The dynamic label encoder includes a boundary extension unit for automatically extending the protection range and updating the label tensor according to the feature type.

Images