Large-angle virtual license plate data set construction method

Abstract

The invention discloses a large-angle virtual license plate data set construction method. The method comprises the steps of 1, controlling and outputting basic license plate texture resources through batch automatic texture generation; 2, executing modeling parameter control and material attribute setting by utilizing a modeling tool and an engine, and generating a three-dimensional license plate model with real physical attributes; 3, simulating imaging effects of actual shooting equipment under different angle conditions, covering large-angle imaging of an extreme pitch angle and a roll angle, and outputting view angle transformation parameters; 4, complex illumination change and background interference in the natural environment are simulated, and the sense of reality and complexity of the final image are improved; and 5, designing a batch automatic data generation engine, and driving and coordinating the operation of the previous steps by the engine to realize large-scale, high-efficiency, diversified and full-automatic control virtual license plate data set construction. The recognition effect of the unmanned aerial vehicle in executing the license plate recognition task is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of virtual license plate simulation, and particularly relates to a large-angle virtual license plate dataset construction method. BACKGROUND

[0002] Existing license plate recognition systems mainly rely on image acquisition devices to recognize license plate images from the front or near front of a vehicle. However, in actual unmanned aerial vehicle license plate detection processes, due to complex environments and changes in shooting angles, especially in large-angle skew, long-distance monitoring, high-altitude monitoring, and unmanned aerial vehicle monitoring scenarios, license plate images often exhibit strong perspective distortion, distortion, blur, partial occlusion, and other problems, resulting in a significant decline in the recognition performance of existing license plate recognition algorithms.

[0003] In order to improve the robustness of the license plate recognition system in complex scenarios, constructing a license plate image dataset containing multiple angles, especially large-angle shooting, is a key link for model training and evaluation. However, most traditional real license plate datasets are concentrated in small angles or ideal shooting conditions, lacking license plate samples under large-angle, complex distortion conditions. Obtaining a large number of real large-angle license plate images not only requires complex shooting platforms, large-scale vehicle samples, multi-angle acquisition control, but also may involve privacy and compliance issues, with high acquisition costs, difficult data annotation, and difficult data balance and diversity.

[0004] The existing virtual license plate data generation methods generally have the following shortcomings: 1. The generation process has limited support for large-angle scenarios, especially in extreme deflection, tilt, overhead view, and downward view, lacking realistic perspective modeling and rendering control; 2. The realism is poor, and it is difficult to effectively simulate real-world complex background information such as lighting changes, blur, occlusion, reflection, pollution, aging, and stains; 3. There is a domain gap between virtual data and real data, and the effect of directly training the model is limited; 4. Lack of scalable automated batch generation framework, low generation efficiency, and insufficient flexibility.

[0005] Therefore, how to propose an efficient large-angle virtual license plate dataset construction method that can comprehensively cover different shooting angles, distances, lighting conditions, and environmental interference, while having high realism and scalability, has become an important technical problem that needs to be solved in the field of license plate recognition training data preparation. SUMMARY

[0006] In order to overcome the defects existing in the prior art, the present application provides a large-angle virtual license plate dataset construction method, which combines virtual environment creation, virtual vehicle body creation, virtual license plate creation and dataset construction method, and can efficiently construct a large-angle virtual license plate dataset, thereby improving the recognition effect of the unmanned aerial vehicle when performing the license plate recognition task.

[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is: A large-angle virtual license plate dataset construction method, comprising the following steps: Step 1: Virtual texture library synthesis: Through the coding rule driven virtual number generation, font rendering and layout, material simulation and texture detail generation, microscopic reflection and highlight modeling, and through batch automatic texture generation control output basic license plate texture resources. The basic license plate texture resources referred to in the present application include character patterns, material texture, optical reflection characteristics and other information that can be directly used for three-dimensional modeling, which is a digital collection of the visual and physical properties of the license plate surface; Step 2: Three-dimensional geometric modeling: Based on the basic license plate texture resources output in step 1, use modeling tools and engines to perform modeling parameter control and material property setting to generate a three-dimensional license plate model with real physical properties; Step 3: View angle control: Receive the three-dimensional license plate model generated in step 2, import the three-dimensional model data containing complete physical properties into a virtual shooting simulation system, which is used to simulate the imaging effect of the actual shooting device of the unmanned aerial vehicle under different angle conditions, cover the large-angle imaging of extreme pitch angle and roll angle, and output the view angle transformation parameter; Step 4: Light and background simulation: Combine the view angle transformation parameter output in step 3 with the three-dimensional license plate model generated in step 2 to simulate the complex light changes and background interference in the natural environment, and improve the realism and complexity of the final image; Step 5: Batch automatic data generation: Design a batch automatic data generation engine, which drives and coordinates the operation of the aforementioned steps 1-4 modules to realize large-scale, efficient, diversified and fully automatic control of virtual license plate dataset construction.

[0008] The engine includes a full-parameter control system, a random control logic, a multi-task scheduling architecture, a parallel acceleration mechanism and an automatic labeling synchronous output module.

[0009] The step 1 is specifically: Step 1-1: Encoding rule-driven virtual number generation, establish a rule library containing multi-country and multi-region license plate encoding specifications; and support automatic generation of legal license plate number combinations according to rules; support proportional balanced sampling of specific license plate categories to ensure sample diversity; Step 1-2: Font rendering and typesetting, call standard font library, use Fangzhong Black, DIN1451, FESchrift fonts for rendering; use image drawing library PIL, OpenCV to process license plate images, draw license plate characters into transparent texture images; support font size, spacing, line spacing perturbation, simulate printing deviation; Step 1-3: Material simulation and texture detail generation, generate different reflective film material effects through the program, simulate high-reflective microbead layer, coating thickness, wear and tear effects; restore license plate details by superimposing surface stains, scratches, oil stains, fading, and oxidation micro-textures; and simulate production batch errors and manufacturing defects; Step 1-4: Microscopic reflection and highlight modeling, calculate reflection effects at different incident angles through the micro-surface BRDF physical model, and use the Gaussian scattering model to finely control high-angle reflection spot changes; Step 1-5: Batch automatic texture generation control, through the batch task scheduling system, realize the automatic generation of ten million textures, and automatically record the generation parameter tags of each texture sample at the same time, which is convenient for data tracing and management.

[0010] The step 2 is specifically: Step 2-1: Modeling tools and engines, modeling is performed through Blender, which is an open-source 3D modeling and rendering software that supports Python API automatic control. All modeling parameters are controlled through a parameterized control interface for batch programmable modeling. The modeled model is exported for subsequent rendering engine calls; Step 2-2: Modeling parameter control, license plate length and width are adjusted according to different national standards; thickness control is simulated from standard thin metal plate to thick plastic plate; frame type modeling is performed for rectangular, rounded rectangular, arc edge, and solid frame; surface curvature control simulates slight bending and warping of the plate during natural use; local subtle concave, plate micro-wavy undulation simulation; Step 2-3: Material property setting, choose different plate metal texture and plastic composite material to simulate; or customize the refractive index, reflectivity, and diffuse reflection ratio parameters of the reflective film; simulate the actual reflective scattering characteristics through micro-surface roughness; realize automatic batch application of mapping through dynamic mapping of texture images generated by the virtual texture library.

[0011] The perspective transformation parameters in step 3 include three-dimensional coordinates of camera coordinate system simulation, camera optical axis orientation, and sensor size parameters, deflection angle, pitch angle, and spin angle parameters covering extreme angle imaging, perspective projection matrix parameters describing imaging distortion.

[0012] Step 3 specifically includes: Step 3-1: Camera coordinate system simulation, defining virtual camera position and orientation in three-dimensional space, dynamically adjusting the camera on the unmanned aerial vehicle in horizontal, pitch, and spin three-dimensional directions, simulating the perspective projection model when the unmanned aerial vehicle actually takes pictures; Wherein, the perspective projection model is realized by the following formula: let the homogeneous coordinates of the three-dimensional space point be: ; After the camera intrinsic matrix , the rotation matrix , and the translation vector are transformed, the homogeneous coordinates on the image plane are obtained, which satisfy ; Wherein, the intrinsic matrix , , is the focal length in the x and y directions, , is the principal point coordinate; the rotation matrix is generated by the horizontal angle , the pitch angle , and the spin angle through the Euler angle conversion formula, that is, ; The translation vector corresponds to the position offset of the virtual camera in the world coordinate system; Step 3-2: Large-angle parameterization control, complete virtual camera coordinate system simulation, dynamically adjust the camera in horizontal deflection angle , pitch angle , and spin angle three spatial degrees of freedom, while supporting custom setting shooting distance (from close-up to high-altitude long-distance view, ), lens focal length (converted to pixel focal length by , , , is the pixel size), field of view angle , and depth of field parameters, wherein , is the sensor parameter, and the depth of field parameter is obtained by the aperture value and focal length Calculations were performed to highly simulate the imaging posture of actual shooting equipment in various complex scenarios; Step 3-3: During the image rendering process, the system calculates and outputs the complete perspective projection matrix parameters in real time. This facilitates subsequent annotation output, model geometry enhancement, and training correction.

[0013] This module is widely applicable to various complex shooting scenarios, such as drone high-altitude upward shooting, vehicle-mounted oblique rear dynamic tracking shooting, overhead monitoring under overpasses, and long-range fixed-point acquisition. It effectively breaks through the limitations of traditional datasets that are mainly based on frontal perspectives, comprehensively covers large-angle acquisition samples in extremely complex scenarios, and improves the breadth of the dataset.

[0014] In step 4, complex lighting changes and background interference include multi-light source modeling, day and night and weather simulation, environmental occlusion simulation, specular reflection simulation, and background material automatic switching library.

[0015] Step 4 specifically involves: Step 4-1: Multi-light source modeling, including simulation of sunlight and street light, as well as simulation of secondary light sources from multiple directions such as ambient diffuse light and building reflected light, and the intensity, direction and color temperature of the light sources can be dynamically adjusted; Step 4-2: Day and night and weather simulation. Classify the environment into time periods including sunny, cloudy, dusk and night. For complex weather conditions, add simulations of rain, snow, fog, frost, etc., or simulate the dynamic mirror effect of wet and slippery road surfaces. Step 4-3: Environmental occlusion simulation. Add dynamic projected shadows from leaves, buildings, traffic lights, and pedestrians to create a realistic environment. These occluders change in real time with the viewing angle to enhance realism. Step 4-4: Mirror reflection simulation, simulating glass reflection, window mirror highlights, interference from reflections from adjacent vehicles, and reflection spots from wet and slippery ground surfaces. Steps 4-5: Establish various scene libraries, simulating multiple scenarios such as city streets, highways, mountain roads, parking lots, tunnels, and toll stations, randomly combining environmental obstructions, weather, and scene libraries. This training set will not be monotonous and will be sufficiently large.

[0016] Step 5 specifically involves: Step 5-1: Fully parametric control system. The batch generation engine integrates all adjustable control parameters of the aforementioned modules (virtual texture library, 3D geometric modeling, viewpoint control, lighting simulation, background occlusion, etc.). Each control parameter is accessible through a configuration file or interface, including the following parameters: License plate code type, font style, and color category; The curvature, thickness, and border style of the 3D geometric model; Viewpoint control: Yaw, Pitch, Roll, distance, focal length; light intensity, light source angle, day / night cycle, weather type; The probability of environmental occlusion and the category of occluded objects; All parameters support setting fixed values, distribution function sampling, or completely randomized sampling modes, and support flexible configuration of training set sampling strategies with different complexities; Step 5-2: Random control logic. When generating batch tasks, the system performs multi-dimensional parameter space joint sampling through a highly controllable random controller. Each sample is an independent random combination of all dimensional parameters to ensure high diversity among samples. It supports setting complexity weight control (such as controlling the increase of high angle and high distortion ratio to enhance model robustness). The random seed mechanism ensures that the generation process of each batch of samples is reproducible, which is convenient for version management and traceability analysis. Step 5-3: Multi-task scheduling architecture. The engine has a built-in multi-task scheduling and management system that supports task splitting and parallel processing, improves task throughput, and can be deployed on multi-core single machines, local multi-node server clusters or cloud computing cluster platforms to achieve high-speed batch automatic continuous generation of tens of millions of samples. It also has dynamic resource monitoring, fault tolerance and recovery, and task breakpoint resume. Step 5-4: Parallel rendering acceleration mechanism. The rendering engine is based on a GPU parallel acceleration architecture and uses CUDA cores to improve the parallel efficiency of image rendering. It integrates the OptiX real-time path tracing acceleration library and uses multi-GPU multi-core parallel rendering, which greatly reduces the image rendering time. Each sample can be generated in just a few seconds. Step 5-5: Automatic annotation and synchronous output module. Upon completion of rendering each image, the system synchronously outputs complete and matching annotation data in real time, including: the coordinates of the four points of the license plate boundary, represented by the two-dimensional pixel coordinates of the four corner points under perspective distortion; perspective projection matrix parameters, fully recording the perspective matrix of the corresponding viewpoint, supporting subsequent geometric correction training; character segmentation coordinate labels, with individual bounding box coordinates for each character, adapting to character segmentation and character recognition tasks; license plate category and scene attribute labels, recording the license plate type of the sample (country / region code type, energy type, special purpose category) and the shooting scene classification (e.g., city, highway, tunnel, parking lot, etc.); distortion intensity labeling information, quantifying and recording distortion disturbance parameters such as motion blur level, lens distortion parameters, noise level, and occlusion coverage ratio; and multi-format annotation output, supporting simultaneous export of annotation standard files in multiple formats such as COCO, YOLO, VOC, and MMDetection.

[0017] The beneficial effects of this invention are: 1. Strong realistic physical modeling capability: Through the view control module in step 3, the camera can be tilted at a horizontal angle. Pitch angle and spin angle The dynamic adjustment within a large angle range, combined with the perspective projection matrix formula for precise calculation, can simulate extreme large-angle view distortion; at the same time, the three-dimensional geometric modeling module in step 2 simulates the physical properties of the license plate such as size, thickness, and curvature, which can realistically restore the spatial distortion and optical deformation under complex imaging conditions. 2. Highly realistic compositing effect: The virtual texture library compositing module in step 1 generates realistic textures through material simulation, microscopic reflection modeling, etc.; the lighting and background simulation module in step 4 performs multi-light source modeling, day and night and weather simulation, environmental occlusion simulation, etc. The multi-level post-processing modules work together to effectively supplement the interference of real distortion and realistically simulate the real visual effect under complex shooting scenes. 3. Full-process parameterized automatic control: The batch automated data generation engine in step 5 includes a full-parameterized control system, which integrates adjustable parameters of each module and supports fixed value, distribution function sampling or completely randomized sampling modes. Combined with random control logic, multi-task scheduling architecture and parallel acceleration mechanism, it realizes batch automatic control generation of complex multi-parameter data, which greatly improves data synthesis efficiency and scene adaptability. 4. Automatic Synchronous Output of Annotation Information: The automatic annotation synchronization output module in step 5-5 outputs high-precision annotation information such as the coordinates of the four points of the license plate boundary, perspective projection matrix parameters, and character segmentation coordinate labels in real time when the image rendering is completed. No manual annotation is required, which can directly improve the availability of training data. 5. High scalability: Step 1-1 supports license plate encoding rules from multiple countries and regions; Step 2 allows adjustment of license plate modeling parameters to adapt to different vehicle categories; Step 3 can simulate shooting parameters of different monitoring equipment; and Step 4 allows setting various lighting conditions. Through the flexible configuration of these modules, complex application needs can be quickly expanded, making it widely applicable. 6. Significantly improved training results: When training with the data generated by this invention, step 3 covers large-angle samples in complex scenarios such as high-altitude and oblique rear view of UAVs, and step 4 simulates various environmental interferences. Combined with accurate automatic annotation information, the recognition accuracy and robustness of the model in complex applications such as high-altitude monitoring, oblique rear view, and UAV scenarios can be significantly improved. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall architecture of the large-angle virtual license plate dataset construction method of the present invention.

[0019] Figure 2 This is a diagram showing the parameter control structure for the three-dimensional physical license plate modeling of the present invention.

[0020] Figure 3 This is a schematic diagram illustrating the effect of the illumination and environmental interference simulation module of the present invention.

[0021] Figure 4 This is the flowchart of batch automated sample generation and annotation output for the present invention. Specific embodiments

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] The present invention provides a method for constructing a large-angle virtual license plate dataset, aiming to provide a method for constructing a large-angle virtual license plate dataset with high realism, multi-view controllability, and large-scale automation capabilities, effectively overcoming the problems of data scarcity, lack of realism, and low synthesis efficiency in existing real data collection and data augmentation methods in complex large-angle scenarios, and providing rich, real, and diverse training samples for the training, testing, and model evaluation of license plate recognition systems. The following is a detailed description of the technical solution: Step 1. Virtual texture library synthesis module To avoid privacy issues, legal restrictions, and insufficient data balance in the collection of real license plate images, a complete license plate texture library is directly constructed using fully virtual generation technology. This module completely eliminates the need for real photo input and directly generates a large-scale, highly realistic virtual license plate texture data relying on procedural design and image synthesis algorithms. The specific operation steps are as follows: Step 1-1: Virtual number generation: A license plate coding specification rule library covering multiple countries and regions is established and stored in JSON format, covering metadata such as character sets (e.g., the list of province abbreviations for Chinese license plates, letter and number combination rules), digit limits (e.g., 7 digits for ordinary license plates, 8 digits for new energy vehicle license plates), and special symbol positions (e.g., the separator ".") in different countries and regions. Random generation is implemented based on the random library in Python: first, the province abbreviation (e.g., "Yue", "Jing") is extracted according to the rule library, then letters are selected through a weighted random algorithm (considering the difference in letter usage frequencies), and finally, a number combination is generated. After generation, the legality is verified through regular expressions (e.g., excluding the appearance of "O" and "I" in specific positions). Balanced sampling is achieved through a preset ratio dictionary (e.g., 15% for new energy vehicle license plates, 5% for police vehicle license plates), and a stratified sampling algorithm is used to ensure that the number of samples of each type meets the expectations.

[0024] Step 1-2: Font rendering and layout: The font file is loaded using the ImageFont module from Python's PIL library, and the characters are rendered using the text() function of the ImageDraw module. The font size parameter is calculated according to the license plate size ratio (e.g., the height of Chinese license plate characters is approximately 90 pixels). A random offset based on a normal distribution (mean 0, standard deviation 0.3 pixels) is added when calculating the character coordinates, and the line spacing deviation is controlled within ±0.5 pixels. Micro-scaling (scaling factor 0.98-1.02) is achieved using OpenCV's warpAffine function.

[0025] Steps 1-3: Material Simulation and Texture Detail Generation The reflective film material uses a Photoshop script to batch generate basic textures, then Python's NumPy library is used to simulate a microbead layer effect, generating a random dot matrix of 50-200 micrometers. Gaussian blur is used to simulate reflective diffusion. Stain generation employs the Perlin noise algorithm: the libnoise library is used to generate a grayscale noise map, which is then thresholded to obtain the stain area. During overlay, the alpha channel (transparency 30%-70%) is adjusted to control the display intensity. Scratch effects are generated using random line segments: line length 5-20mm, width 1-3 pixels, grayscale value 100-200 (simulating different degrees of wear), drawn using OpenCV's line() function, with randomized orientation.

[0026] Steps 1-4: Microscopic Reflection and Specular Modeling The BRDF model employs an improved Blinn-Phong model, calculating reflected light intensity using ambient light, diffuse reflection, and specular components. The Gaussian scattering model is achieved by generating a 2D Gaussian distribution matrix: the specular region has the highest brightness at its center, expanding outwards towards the edges... attenuation, The value increases as the angle of incidence increases.

[0027] Steps 1-5: Batch Automated Texture Generation Control A batch task scheduling system based on Celery distributed task queues is implemented. The texture generation function is encapsulated as a task, and the parameters in steps 1-1 to 1-4 above are set as configurable items. Redis is used as a message broker to distribute tasks to multiple worker nodes. Parameter logs are recorded in CSV format, including sample ID, generation time, and all adjustable parameter values ​​(such as font name, stain type, etc.). Data aggregation and filtering (such as filtering samples with "oil stains + heavy wear") are implemented using the pandas library.

[0028] Step 2. 3D Geometric Modeling Module: A standard license plate 3D physical model is created using 3D modeling tools. The following specific steps are followed to achieve parameter adjustment and simulation of realistic physical properties: Step 2-1: Modeling Tools and Engine Initialization: Using Blender 3.0+, a modeling script was written using its Python API (bpy module). A planar object was created as the license plate model, and its default size was set to 440mm × 140mm according to the Chinese license plate standard.

[0029] Step 2-2: Surface Modeling Parameter Control: The size of the license plate surface is adjusted by modifying the `dimensions` property of the mesh. Custom dimensions are input via a JSON configuration file; the script parses the file, dynamically modifies the `dimensions` parameters, and updates the UV mapping coordinates to prevent texture stretching. Thickness is obtained by adding a solidification modifier to the license plate plane; the input thickness parameter is directly mapped to the `modifier.thickness` property. The license plate border is implemented using Boolean operations: first, create an outer frame 2mm larger than the main body, then subtract the main body area using a difference operation. Rounded corners are achieved by adding a chamfer modifier, with the `segments` parameter controlling the smoothness of the rounded corners. Warping and indentations on the license plate surface are achieved using a displacement modifier, generating a random grayscale image as the displacement texture; the intensity parameter controls the indentation depth.

[0030] Steps 2-3: Material Property Control: Create a PBR material in Blender's material editor and set node parameters using a Python script. Different license plate materials correspond to different Metallic and Roughness parameters. The refractive index of the license plate surface is controlled by the IOR parameter, and the diffuse reflection ratio is adjusted by the alpha value of the BaseColor node in the material.

[0031] Step 3. Large-angle viewing angle control module To achieve precise control of viewpoint parameters and large-angle scene simulation in a 3D scene, follow these steps: Step 3-1: Camera coordinate system simulation: Create a CineCameraActor as a virtual camera in UE5. Set the initial values ​​of the intrinsic parameter matrix through lens properties: Focal Length is set to 50mm, and the principal point coordinates (FilmbackOrigin) are set to (18mm, 12mm). Obtain the intrinsic parameter matrix K in real time through the Blueprint node "GetCameraIntrinsics". Set the camera's AttachParent to the license plate model, and set the camera's relative position to (0,0,0) through "SetActorRelativeLocation" to ensure that the rotation revolves around the center of the license plate.

[0032] Step 3-2: Large-Angle Parametric Control: Euler angle transformation is implemented through the Actor Transform component and dynamically adjusted via the Blueprint node "AddActorWorldRotation". Shooting distance adjustment is achieved by translating along the camera's forward vector: obtain the camera's ForwardVector in the Blueprint, multiply it by the distance parameter, and update the position using "SetActorWorldLocation". Focal length adjustment is achieved by directly modifying the CineCamera's FocalLength property.

[0033] Step 3-3: Real-time output of perspective projection matrix: As described in the formula above, after obtaining the Filmback property and Euler angles through UE5, the intrinsic parameter matrix K and rotation matrix R can be calculated. The perspective matrix M is calculated through the Blueprint node "MultiplyMatrixbyMatrix" and is attached to the metadata output each frame during rendering through "RenderTarget".

[0034] Step 4. Lighting and Environment Background Modeling Module Integrate complex lighting and background simulation in the UE5 rendering engine through the following steps: Step 4-1: Multi-Light Source Modeling: The main light source uses DirectionalLight, with "RayTracedShadows" enabled and Intensity set to 100,000 lux (sunny day). The light baking quality is improved using "LightmassImportanceVolume". Streetlights use PointLightwithShadows, with AttenuationRadius set to 20m, Intensity to 50,000 lux, and Temperature to 2700K. The auxiliary light source uses SkyLight, with "RealTimeCapture" enabled and Intensity set to 15% of the main light source. This simulates possible ambient diffuse reflection and allows adjustment of the reflection hue via "LightColor".

[0035] Step 4-2: Day / Night and Weather Effects Simulation: Simulate changes in weather and the alternation of day and night by adjusting the intensity of the main light source and the color of the auxiliary light source as described in the previous steps. Fog effects are achieved using the "VolumetricFog" component: FogDensity is set to 0.02-0.1, FogAlbedo is set to (0.9, 0.9, 0.9), and "HeightFalloff" is enabled to simulate dense fog near the ground. Rain effects are achieved using the "NiagaraParticleSystem," with an emissivity of 1000 particles / second, a lifespan of 1-2 seconds, a velocity of 15 m / s, and a rendering mode of "Additive." Modifying the particle emissivity allows for simulation of different weather conditions.

[0036] Step 4-3: Ambient Occlusion and Shadow Casting: Occlusion objects are created as BlueprintActors by category. Instantiation and recycling are managed through the "ActorPool" plugin. "SpawnActorfromClass" is called to randomly generate occluders within a 1-10m range around the license plate. Collision is detected using "LineTracebyChannel" before generation. After generating occluders, the quality and smoothness of the generated shadows can be controlled by adjusting the ShadowMapResolution and ShadowFilterSharpness of DirectionalLight.

[0037] Step 4-4: Automatic Scene Switching and Blending: UE5's "WorldPartition" manages large scenes, including sub-levels such as cities, highways, parking lots, and mountainous areas. Scenes are dynamically loaded using "LoadLevelInstance," with loading priority set according to scene probability. Perspective matching is achieved through "CameraAnimation": Multiple camera positions are preset in the background level, and "MatchCameraatoViewTarget" matches the current camera parameters with the camera positions to ensure consistent background perspective with the license plate view. Dynamic switching is achieved through "LevelStreaming": Every 1000 samples generated, "UnloadLevelInstance" is called to unload the current background and randomly load a new background level.

[0038] Step 5. Batch Automated Data Generation Engine The following steps will enable the automatic generation engine of large-scale, high-efficiency virtual license plate datasets: Step 5-1: Building a Fully Parameterized Control System: Parameter configuration is defined using JSON Schema, including metadata such as parameter name, type, range, default value, and sampling method. The configuration file is parsed using Python's ConfigParser module, mapping parameters to API calls of each module and supporting parameter dependency validation.

[0039] Step 5-2: Implementation of random control logic: Pass in a random seed through command line parameters to achieve reproducible random sample generation.

[0040] Step 5-3: Multi-task architecture implementation: A multi-task architecture is implemented using the Celery framework in Python: Task producers receive generation requests and send them to a Redis queue; worker nodes subscribe to the queue and call rendering functions to execute tasks. Load balancing is achieved through Celery's autoscaler: automatically adjusting the number of worker nodes based on the queue length and monitoring the CPU utilization of each node. This architecture can simultaneously handle task timeout resending and breakpoint resumption exceptions.

[0041] Step 5-4: GPU Parallel Accelerated Rendering: Using the batch rendering function in UE5Sequencer, sample generation tasks are allocated frame by frame. Command-line rendering is executed via "Commandlet," reducing the rendering time for a single 1920×1080 image to less than 0.5 seconds. The hardware execution layer uses RenderTargetBatchProcessing for parallel rendering. By creating multiple RenderTargets, the GPU's multi-stream processors can handle multiple rendering instances simultaneously.

[0042] Step 5-5: Automatic annotation and synchronous output: Annotated data is temporarily stored in UE5's "DataTable". Every 100 samples generated, a Python script is called to export them in COCO / YOLO format. The script accesses scene data through "UnrealPythonAPI" to ensure that the annotations are synchronized with the images.

[0043] The batch engine can efficiently generate millions of high-quality synthetic samples, adapting to multi-task training scenarios such as license plate recognition detection, localization, segmentation, recognition, and correction.

[0044] Figure 1The overall structure of this invention is illustrated. The method for constructing a large-angle virtual license plate dataset comprises five core modules. Arrows indicate the data flow of each module, presenting the entire process logic from virtual texture generation to batch data output. The bottom layer is the license plate texture library synthesis module (step 1), which serves as the foundation for constructing the dataset, outputting virtual license plate texture resources containing valid numbers and realistic material textures. The second layer is the 3D physical modeling module (step 2), which, after receiving the texture resources, constructs 3D license plate models with different physical properties through parametric control, transforming planar textures into 3D solid models. The third layer is the viewpoint control module (step 3). Through camera coordinate system simulation and large-angle parametric control, perspective transformation is applied to the 3D license plate model to restore the license plate distortion characteristics at extreme angles. The fourth layer is the lighting and background simulation module (step 4), which, based on the model after viewpoint transformation, superimposes real-world environmental interference such as multiple light sources, weather, occlusion, and reflections to enhance image realism. The top layer is the batch automated data generation engine (step 5), which coordinates the batch operation of the first four modules through parametric control, random logic, and parallel acceleration.

[0045] Figure 2 This paper illustrates the 3D physical license plate modeling parameters and their hierarchical relationship of this invention. The figure centers on the 3D license plate model, with branching structures indicating the categories of parameters. Basic size parameters include license plate length, width, and thickness, which can be adjusted according to different national or regional standards; border parameters include chamfer curvature, border width, and border thickness, implemented through a chamfer modifier and Boolean operations; surface morphology parameters include curvature and random indentation depth, simulating deformation during natural use; material parameters include texture mapping (pointing to the texture resources and surface materials in step 1), and support adjustment of surface metallicity and reflectivity parameters.

[0046] Figure 3 This diagram visually demonstrates the functionality and technical effects of the lighting and background simulation module of this invention. Using the same license plate model at the same angle as a baseline, the diagram shows the simulation effects under multiple scene interference conditions in four stages from left to right. The multi-source modeling module simulates the basic lighting environment, such as natural light on a sunny day, without additional environmental interference. The day / night and weather simulation module superimposes day / night cycle changes and weather interference onto the basic lighting, simulating the challenges of license plate recognition under low light and extreme weather conditions. The environmental occlusion simulation module introduces occlusion objects, randomly obscuring local areas of the license plate, recreating the complex situation of natural object occlusion in road scenes. The specular reflection simulation module simulates reflected light from a smooth surface, causing optical interference such as light spots and mirror images on the license plate surface.

[0047] Figure 4The process of batch automated data generation and annotation of the present invention is demonstrated. The controller receives a parameterized configuration file input by the user (including parameters such as the number of samples, license plate type, angle range, lighting conditions, and sampling method), schedules each working node to generate virtual license plate images, and generates automatic annotation information, including the pixel coordinates of the four corner points of the license plate under perspective distortion, perspective projection matrix, license plate character information, and scene attribute labels.

Claims

1. A method for constructing a large-angle virtual license plate dataset, characterized in that, Comprising the following steps; Step 1: Control the output of basic license plate texture resources through batch automatic texture generation by encoding rule-driven virtual number generation, font rendering and typesetting, material simulation and texture detail generation, microscopic reflection and highlight modeling; Basic license plate texture resources include information about character patterns, material texture, and optical reflection characteristics directly used for three-dimensional modeling; Step 2: Based on the basic license plate texture resources output in step 1, use modeling tools and engines to perform modeling parameter control and material property setting to generate a three-dimensional license plate model with real physical properties; Step 3: Receive the three-dimensional license plate model generated in step 2, import the three-dimensional model data containing complete physical properties into the virtual shooting simulation system, simulate the imaging effect of the actual shooting device of the unmanned aerial vehicle under different angle conditions, cover the large-angle imaging of the extreme pitch angle and roll angle, and output the perspective transformation parameter; Step 4: Combine the perspective transformation parameter output in step 3 to act on the three-dimensional license plate model generated in step 2 to simulate complex light changes and background interference in the natural environment, and improve the realism and complexity of the final image; Step 5: Design a batch automatic data generation engine that drives and coordinates the operation of the preceding steps to achieve virtual license plate dataset construction.

2. The method of claim 1, wherein, The step 1 is specifically: Step 1-1: Establish a license plate coding specification rule library containing multiple countries and regions; and support automatic generation of legal license plate number combinations according to the rules; support proportional balanced sampling of specific license plate categories to ensure sample diversity; Step 1-2: Call the standard font library, use Fangzhong Black, DIN1451, and FESchrift fonts for rendering; use the image drawing library PIL and OpenCV to process license plate images, and draw license plate characters into transparent texture images; support font size, spacing, and line spacing perturbation to simulate printing deviations; Step 1-3: Generate different reflective film material effects through the program to simulate the effects of high-reflective microbead layer, coating thickness, and wear and tear; restore license plate details by superimposing surface stains, scratches, oil stains, fading, and oxidation micro-textures; and simulate production batch errors and manufacturing defects; Step 1-4: Calculate the reflection effect at different incident angles through the micro-surface BRDF physical model, and use the Gaussian scattering model to finely control the high-angle reflection spot variation; Step 1-5: Through the batch task scheduling system, realize the automatic generation of ten million textures, and automatically record the generation parameter tags of each texture sample at the same time to facilitate data tracing and management.

3. The method of claim 2, wherein, The step 2 is specifically: Step 2-1: Model through Blender, all modeling parameters are controlled through a parameterized control interface for batch programmable modeling, and the modeled model is exported for subsequent rendering engine calls; Step 2-2: Adjust the length and width of the license plate according to different national standards; simulate from standard thin metal plates to thick plastic plates for thickness control; model straight rectangular, rounded rectangular, arc edge, and solid frame for frame type; simulate slight bending and warping of the plate surface during natural use for surface bending control; simulate local subtle concave, plate micro-wavy undulation; Step 2-3: Choose different plate metal texture and plastic composite material to simulate; Or customize the refractive index, reflectivity, and diffuse reflection ratio parameters of the reflective film; Simulate the actual light scattering characteristics through micro-surface roughness; Realize automatic batch application of mapping through the texture images generated by dynamically mapping the virtual texture library.

4. The method of claim 3, wherein, In step 3, the view transformation parameters include the three-dimensional coordinates of the camera coordinate system simulation, the camera optical axis orientation and the sensor size parameters, the deflection angle, the pitch angle and the spin angle parameters covering the extreme angle imaging, and the perspective projection matrix parameters describing the imaging distortion.

5. The method of claim 4, wherein, The step 3 is specifically: Step 3-1: Define the virtual camera position and orientation in three-dimensional space, dynamically adjust the camera in horizontal, pitch and spin three-dimensional directions, and simulate the perspective projection model when the unmanned aerial vehicle actually shoots; Wherein, the perspective projection model is realized by the following formula: let the homogeneous coordinates of the three-dimensional space point be: ； through the camera intrinsic matrix , a rotation matrix , and a translation vector After the transformation, the homogeneous coordinates on the image plane are obtained , which satisfy ; where the intrinsic matrix , , is the focal length in x and y directions, , is the principal point coordinates; the rotation matrix is generated by the horizontal angle , the pitch angle , the roll angle through the Euler angle conversion formula, i.e. ； Translation vector a corresponding virtual camera position offset in the world coordinate system; Step 3-2: Complete the virtual camera coordinate system simulation, the camera dynamically adjusts on three spatial degrees of freedom of horizontal deflection angle , pitch angle and spin angle , while supporting custom settings for shooting distance (from close-up to high-altitude long-distance view, ), lens focal length (converted to pixel focal length, , pixel size, , field of view angle and depth of field parameters, where , are sensor parameters, and the depth of field parameters are calculated by the aperture value and focal length to highly simulate the imaging posture of actual shooting equipment in various complex scenes; Step 3-3: During the image rendering process, the system calculates and outputs the full perspective projection matrix parameters in real-time for subsequent annotation output, model geometry enhancement, and training correction usage.

6. The method of claim 5, wherein, In step 4, the complex light change and background interference include multi-light source modeling, day and night and weather simulation, environment occlusion simulation, mirror reflection simulation, and background material automatic switching library.

7. The method of claim 6, wherein, The step 4 is specifically: Step 4-1: Multi-light source modeling, light sources include simulation of sunlight and street light, as well as simulation of environment diffuse light and building reflection light and other multi-direction secondary light sources, and light intensity, direction and color temperature can be dynamically adjusted; Step 4-2: Classify the environment into time period changes such as sunny, overcast, dusk and night, and add environment simulation such as rain, snow, haze and frost for complex weather conditions, or simulate environment wet road surface high reflection dynamic mirror effect; Step 4-3: Add dynamic projection shadow of tree leaves, buildings, traffic signal lights and pedestrians, and establish a real environment; Step 4-4: Mirror reflection simulation, simulate glass reflection, car window mirror high light, adjacent vehicle reflection interference and ground wet water reflection spot simulation; Step 4-5: Establish various scene libraries to simulate city streets, highways, mountain roads, parking lots, tunnels and toll stations, and randomly combine environment occlusion, weather and scene library.

8. The method of claim 7, wherein, The step 5 is specifically: Step 5-1: Full-parameter control system, batch generation engine integrates all adjustable control parameters of the preceding modules; Each control parameter is open through a configuration file or an interface, including the following parameters: License plate encoding type, font style, color category; Three-dimensional geometric model curvature, thickness, border style; Yaw, pitch, roll, distance, focal length of view control; Light intensity, light angle, day and night cycle, weather type; Environment occlusion probability and occlusion object category; All parameters support setting fixed value, distribution function sampling or completely randomized sampling mode, and support flexible configuration of different complexity training set sampling strategies; Step 5-2: During batch task generation, the system uses a highly controllable random controller to perform joint sampling in a multi-dimensional parameter space. Each sample is an independent random combination in all dimensions, ensuring high diversity among samples. The system supports setting complexity weight controls, and the random seed mechanism ensures that each batch sample generation process can be reproduced, facilitating version management and traceability analysis. Step 5-3: Multi-task scheduling architecture, with an engine built-in multi-task scheduling management system, supporting task splitting and parallel processing to improve task throughput. It can be deployed on multi-core single machines, local multi-node server clusters, or cloud computing cluster platforms, achieving high-speed batch automatic continuous generation of millions of samples. It also supports dynamic resource monitoring, abnormal fault recovery, and task breakpoint resume. Step 5-4: Rendering engine based on GPU parallel acceleration architecture, using CUDA cores to improve image rendering parallel efficiency, integrating OptiX real-time path tracing acceleration library, and using multi-GPU multi-core parallel rendering to significantly shorten image rendering time, with each sample generation taking only seconds. Step 5-5: Automatic labeling synchronous output module, which outputs complete supporting labeling data in real time when each image is rendered, including license plate boundary four-point coordinates, represented as four corner two-dimensional pixel coordinates under perspective distortion; perspective projection matrix parameters, which record the corresponding perspective matrix of the view angle, supporting subsequent geometric correction training; character segmentation coordinate labels, each character has its own bounding box coordinates, suitable for character segmentation and character recognition tasks; license plate category and scene attribute labels, recording the license plate type and shooting scene classification of the sample; distortion intensity label information, quantitatively recording motion blur level, lens distortion parameters, noise level, and occlusion coverage ratio distortion disturbance parameters; multi-format labeling output, supporting simultaneous export of COCO, YOLO, VOC, MMDetection, and other multi-format labeling standard files. ​

Images