Multi-dimensional decoupling traffic scene data generation method and device for automatic driving training

Abstract

The invention relates to the field of automatic driving training, and discloses a multi-dimensional decoupling traffic scene data generation method and device for automatic driving training. The method comprises the steps that a user defines a high-order configuration file in advance, and the high-order configuration file is a plurality of scene description files generated based on a combined strategy engine; starting an automatic driving simulation platform service, establishing simulation environment connection, loading a high-precision geographic information model and starting a global synchronization mode; according to the embodiment vehicle, a multi-view perception system is deployed under a coordinate system of an embodiment vehicle body; the environmental factors are dynamically controlled through a time sequence interpolation engine, continuous and smooth evolution of the environmental factors is achieved, and a deterministic synchronous barrier mechanism is used; and collecting data. According to the invention, large-scale and multi-dimensional traffic scenes are automatically generated; carrying out algorithm robustness and generalization verification in a complex environment; high time-space synchronization simulation of the traffic environment is supported; and the scalable and repeatable scene generation capability is realized.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving training technology, and more specifically, to a method and apparatus for generating multi-dimensional decoupled traffic scene data for autonomous driving training. Background Technology

[0002] With the development of autonomous driving, Intelligent Transportation Systems (ITS), and digital twins of urban traffic, the industry's demand for high-quality, highly controllable, and large-scale traffic image data is constantly growing. Currently, traffic data commonly used in research and development mainly comes from two sources: data collected from actual roads and publicly available traffic datasets. However, data collected from actual roads typically has the following shortcomings: long data acquisition cycles, constraints from environmental factors such as weather, lighting, and traffic flow, and a lack of controllability; limited scene coverage, making it difficult to systematically cover multi-dimensional factors; and complex coupling of traffic states, making it difficult to conduct targeted experimental designs. On the other hand, while publicly available traffic datasets provide convenience for general research, they often suffer from insufficient matching with specific business needs and weak correlation, failing to meet the high-precision experimental data requirements of cutting-edge research such as decoupled learning, causal inference, and single-factor traffic behavior modeling.

[0003] Existing technologies still have the following shortcomings in the acquisition and construction of traffic scenario data: First, actual road data collection methods are limited by the uncontrollability of the environment and scenario, resulting in long data collection cycles, high costs, and difficulty in achieving systematic coverage of multi-dimensional factors and targeted experimental design; Second, although publicly available traffic datasets have alleviated the data shortage problem to some extent, they generally suffer from insufficient business relevance and weak scenario customization. In particular, in cutting-edge research such as decoupled learning, causal inference, and single-factor modeling of traffic behavior, they cannot provide high-quality and highly controllable experimental data support, which in turn restricts the research and development and application of algorithms in fields such as autonomous driving, intelligent transportation, and traffic digital twins. Summary of the Invention

[0004] To address at least one of the aforementioned problems, this invention provides a method for generating multidimensional decoupled traffic scene data for autonomous driving training, comprising the steps of: a user pre-defining a high-level configuration file, wherein the high-level configuration file generates multiple scene description files based on a combination strategy engine, and the scene description files define multiple environmental factors in a structured manner, including at least one of map, weather, illumination, time, vehicle speed, vehicle type, vehicle color, and traffic density;

[0005] By initializing the digital twin simulation environment, starting the autonomous driving simulation platform service, establishing a simulation environment connection, loading a high-precision geographic information model and enabling global synchronization mode, loading and parsing the scene description file and obtaining the corresponding high-precision road and topology data;

[0006] Based on the configuration of the scene description file obtained by loading and parsing, a vehicle model of a specified type and color is loaded from the blueprint library of the simulation platform and the vehicle is instantiated. The vehicle position is instantiated through a random point filtering algorithm, and a multi-view perception system is deployed in the coordinate system of the instantiated vehicle body.

[0007] The environmental factors are dynamically controlled by a time series interpolation engine to achieve continuous and smooth evolution of the environmental factors; a deterministic synchronization barrier mechanism is used to synchronously collect data from the multi-view perception system.

[0008] Data from the multi-view perception system is collected, and the collected data is automatically saved to generate a structured, traceable traffic scene dataset.

[0009] Optionally, the step of the combined strategy engine generating multiple scene description files includes: locating several discrete levels for each environmental factor;

[0010] Based on business requirements, multiple scenario description files are automatically generated by manually defining or orthogonal arrays, and each scenario description file corresponds to a test case.

[0011] The scenario description files are batch-scheduled to achieve systematic coverage of the environmental factors.

[0012] Optionally, the time series interpolation engine steps include:

[0013] In each simulation time step, the instantaneous value of the parameter is calculated based on the current simulation timestamp and the preset function type;

[0014] The instantaneous value is updated in real time to the scene environment controller or vehicle controller through the simulation platform's API, achieving the effect of continuous and gradual change of the environmental factors.

[0015] Optionally, the deterministic synchronization barrier mechanism is as follows: after a fixed time step in the simulation calculation, the client enters a blocking waiting state until it confirms that it has received a complete set of multi-view data from the same calculation before it is released;

[0016] Built-in timeout protection logic prevents the system from being permanently blocked due to the loss of data from a single sensor;

[0017] Multi-view image data are aligned with zero deviation at the time of acquisition.

[0018] Optionally, the multi-view perception system includes five virtual cameras deployed in the vehicle's coordinate system: front-view, rear-view, left-view, right-view, and top-view. The virtual cameras maintain spatial consistency with the vehicle body through rigid body transformation and support synchronous acquisition of multimodal data.

[0019] Optionally, the scenario description file is defined in JSON format, supporting static configuration or dynamic functional parameter description, thereby achieving configurability, scalability, and reproducibility of the scenario.

[0020] Optionally, the method further includes the steps of: parsing and scheduling batches of the scene description files through an automated data orchestration and management engine, generating a hierarchical storage directory based on the environmental factor signature, and realizing automated management and data traceability of cross-scene assets.

[0021] Optionally, the steps of dynamic control of the environmental factors include: using a linear interpolation function to achieve continuous changes in parameters such as weather, illumination, and vehicle speed;

[0022] It supports multiple motion modes, including constant speed, constant acceleration, and variable acceleration;

[0023] By applying parameter changes in real time through the simulation API, the gradual process of the real physical world can be simulated.

[0024] Optionally, the process also includes the following steps: automatically destroying the simulation object after each scenario is completed, releasing computing resources, retrieving the next scenario description file from the task queue, seamlessly switching and repeating the execution, achieving unattended batch data generation, and completing resource recycling.

[0025] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0026] 1. Employing a multi-factor decoupled modeling and combination generation mechanism, key factors in traffic scenarios (such as map regions, vehicle types, weather conditions, speed patterns, and spatial layout) are independently described. The factor combination engine automatically instantiates scenarios, enabling rapid construction of large-scale scenarios. The combination engine supports various combination strategies, including full combination generation, random sampling, and orthogonal experimental design. Full combination is used to construct comprehensive and massive training data; random sampling is used to reduce computational load while maintaining diversity; and orthogonal experimental design is used to cover the main factor effects with fewer combinations when the factor space is large and resources are limited, thereby improving generation efficiency. Through these mechanisms, highly scalable scenario generation capabilities are achieved, significantly shortening the preparation cycle required by traditional manual configuration methods, greatly improving data construction efficiency, and reducing overall costs.

[0027] 2. Robustness and generalization verification of algorithms in complex environments: By generating complex scenarios such as continuously changing weather and dynamic speed changes, this invention can provide test samples for perception algorithms under extreme conditions. The generated data can cover extreme and boundary conditions, and support the evaluation of the robustness and generalization ability of autonomous driving perception and decision-making algorithms under complex conditions such as low visibility, sudden acceleration and deceleration, and sudden changes in lighting, thereby improving the reliability of the algorithms in real traffic environments.

[0028] 3. Support for high spatiotemporal synchronization simulation of traffic environment: The system designed in this invention adopts a unified global clock and buffer synchronization mechanism, combined with a timing alignment algorithm, to achieve millisecond-level synchronization of multiple cameras, sensors and simulation control signals. The synchronization error between image frames can be stably controlled within ±10 ms, effectively avoiding multi-view data mismatch and greatly improving the data quality for multimodal perception and timing prediction tasks.

[0029] 4. Scalable and repeatable scene generation capability: All environmental factors and combination strategies are exposed through parameterized interfaces, allowing for the reproduction of any scene as needed, ensuring the repeatability of experiments; at the same time, it supports weight-driven random combinations, which can flexibly generate higher-dimensional and richer traffic scenes, suitable for large-scale training, validation and regression testing, providing high-quality and highly controllable experimental data support, breaking through the constraints of algorithm research and development and application in the fields of autonomous driving, intelligent transportation and traffic digital twins.

[0030] In addition, the present invention provides a multi-dimensional decoupled traffic scene device, including a memory storing a computer program and a processor. When the processor executes the computer program, it implements the multi-dimensional decoupled traffic scene data generation method for autonomous driving training as described above. Attached Figure Description

[0031] Figure 1 The process in the embodiments of the present invention Figure 1 ;

[0032] Figure 2 This is a flowchart of the time series interpolation engine in an embodiment of the present invention;

[0033] Figure 3 This is a flowchart of the deterministic synchronization barrier mechanism in an embodiment of the present invention;

[0034] Figure 4 The process in the embodiments of the present invention Figure 2 ;

[0035] Figure 5 This is a schematic diagram of a pre-set continuously changing function model in an embodiment of the present invention. Detailed Implementation

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following description is provided in conjunction with the accompanying drawings. Figure 1-5 This application will be described in further detail.

[0037] In a first aspect, embodiments of the present invention provide a method for generating multi-dimensional decoupled traffic scene data for autonomous driving training, referring to... Figures 1 to 4The multi-dimensional decoupled traffic scene data generation method for autonomous driving training includes the following steps:

[0038] Step S100: Scenario definition and task queue generation.

[0039] This step aims to define the basic elements required to generate the scene and their patterns of change.

[0040] Step S110: After the program starts, the user predefines a high-level configuration file. The high-level configuration file generates multiple scene description files based on the combination strategy engine, and the multiple scene description files form a task queue.

[0041] The combinatorial engine supports various combinatorial strategies, including full combinatorial generation, random sampling, and orthogonal experimental design. Full combinatorial generation is used to construct comprehensive, massive training data; random sampling reduces computational cost while maintaining diversity; and orthogonal experimental design improves generation efficiency by covering major factor effects with fewer combinations when the factor space is large and resources are limited. Users can choose different engine design scenario description files according to their needs.

[0042] The scene description file defines multiple environmental factors in a structured manner, covering the multidimensional environmental factors required to construct the virtual environment. These environmental factors include at least the following dimensions: 1. Static environment dimension: such as town or rural maps, road topology, etc.; 2. Meteorological and lighting dimension: including weather type (sunny, rainy, foggy, snow, etc.), light intensity, time period (daytime, dusk, night), and solar azimuth angle, etc.; 3. Traffic participant dimension: including vehicle type (cars, trucks, etc.), vehicle color, vehicle speed mode (constant speed, acceleration / deceleration), and traffic flow density, etc. Furthermore, to simulate the dynamic changes of the environment in the real world, this invention introduces a parameterized function definition mechanism. In the configuration file, the above parameters can be not only fixed values ​​but also defined as time functions that change over time. For example, the "fog density (fog_density)" parameter can be defined as a linear interpolation function: "fog_density": {"function_type": "LINEAR_INTERPOLATION", "start_time": 0.0, "end_time": 5.0, "start_value": 0.0, "end_value": 1.0}. This means that the fog density will gradually change from no fog to dense fog within 0 to 5 seconds.

[0043] After the program starts, the main program calls the combined scheduler to process the high-level configuration file and generate multiple independent scene description files. The scene description files adopt JSON format and contain the static configuration or dynamic functional parameter description required for the specific scene, thereby realizing the configurability, scalability and reproducibility of the scene.

[0044] Step S120: Parse and schedule batch scene description files through the automated data orchestration and management engine; the automated data orchestration and management engine can design a structured storage strategy based on factor signatures, generate a hierarchical storage directory based on environmental factor signatures, realize automated management and data traceability of cross-scene assets, realize systematic coverage of environmental factors, and form an unattended data production pipeline.

[0045] Step S200: Simulation environment initialization and scene loading.

[0046] Step S210: Initialize the digital twin simulation environment through the digital twin environment initialization module. The main program (data orchestration engine) first reads a scene description file (JSON) from the task queue. Then, it starts the autonomous driving simulation platform service, establishes a communication channel with the simulation engine, establishes a connection to the simulation environment, and sets a timeout retry mechanism (maximum of 5 attempts). The digital twin environment initialization module has a built-in state verification mechanism to ensure the robustness of the connection.

[0047] Step S220: Force the global synchronization mode to be enabled. Set a fixed time step Δt according to the fields in the above scenario description file, where fps can be configured to 10-30. Lock the update of the simulation world at a fixed time step to provide a basis for the deterministic timing synchronization of all subsequent modules.

[0048] Step S230: Load the high-precision geographic information model and call the road network parsing algorithm to automatically identify and filter out the effective intelligent agent generation location set; the system loads and parses the structured scene description file and obtains the corresponding high-precision road and topology data for subsequent path planning and vehicle navigation.

[0049] Step S240: Call the map API (programming interface) to obtain all available vehicle generation points, and use a validity filtering algorithm to remove points in non-road areas, overlapping obstacle points, and points with insufficient safety distance, to ensure that the distribution of generation points is reasonable and the vehicle generation success rate is ≥99%.

[0050] Step S250: Create a hierarchical storage directory structure based on the metadata in the file: root directory / town name / vehicle type / factor type / location number / viewpoint / frames.

[0051] Step S300: Instantiate the intelligent agent vehicle.

[0052] Step S310: Based on the configuration of the scene description file obtained from the loading and parsing, load the vehicle model of the specified type and color from the blueprint library of the simulation platform and instantiate the vehicle to obtain the 3D model of the vehicle body.

[0053] Step S320: Instantiate the vehicle location using a random point filtering algorithm. That is, randomly select a location from the valid vehicle generation points for actual instantiation. If generation fails, automatically reselect a nearby candidate point and retry until generation is successful.

[0054] Step S330: Configure a PID-based motion controller for the vehicle. The speed factor is recorded in the form of Velocity(profile_type, start_velocity, end_velocity, duration) in the file, and the execution layer generates the speed curve based on this.

[0055] Step S400: Deployment of the multi-view perception system.

[0056] Step S410: Deploy a multi-view perception system in the coordinate system of the already implemented vehicle body; the multi-view perception system includes five virtual cameras deployed in the vehicle body coordinate system according to the parameters preset in the document, namely front view, rear view, left view, right view and top view, and the camera coordinates, attitude and field of view (FOV) are precisely set to ensure coverage of 360° surround view and vertical overhead view.

[0057] Step S420: Enable the camera and vehicle coordinate system binding mechanism. The virtual camera maintains spatial consistency with the vehicle body through rigid body transformation (ensuring that the camera remains strictly synchronized when the vehicle moves and turns), avoiding jitter and parallax.

[0058] Step S430: Configure the image acquisition frame rate and resolution to be determined by the parameter factors in the description file, and support multimodal data synchronous acquisition.

[0059] Step S500: Dynamic control of environmental factors.

[0060] Step S510: In the scene description file, environmental factors (such as weather or speed) are defined as a time function object, rather than a set of static values. This object explicitly describes how the parameters evolve over time, for example: "fog_density": {"function_type": "LINEAR_INTERPOLATION", "start_time":0.0,"end_time":10.0,"start_value":0.0, "end_value":0.8},"velocity":{"profile_type": "LINEAR_ACCELERATION","duration":10.0,"start_velocity": 5.0,"end_velocity": 25.0}.

[0061] Word translation: Weather (precipitation, cloudiness, fog density, sun azimuth angle)

[0062] Wether(): Weather factor configuration(); precipitation: precipitation intensity 0–10, controls the amount of raindrops, rain streaks, watermarks and ground wet reflection effects; cloudiness: cloud cover percentage, 0–10, the higher the value, the darker the sky, the less sharp the sunlight, and the lighter the shadows; fog_density: fog density, range 0–10, the higher the value, the lower the visibility, the blurry distant objects and the reduced contrast; sun_azimuth_angle: the horizontal position of the sun (0–360°), determines the direction of light incidence, and affects the direction of shadows and the brightness distribution of objects.

[0063] Velocity(profile_type,start_velocity,end_velocity,duration)

[0064] Velocity(): Vehicle speed factor configuration(); profile_type: Speed ​​change curve type, defining the shape model of speed growth over time; start_velocity: Initial speed, determining the vehicle's motion state when entering the scene; end_velocity: Final speed, determining the vehicle's target state, such as the end speed of an acceleration / deceleration task; duration: Time required from start to end, controlling the smoothness of the curve change and the speed of the movement. (Note: The Velocity factor is not a fixed number, but a function curve controlled by time).

[0065] Vehicle(asset_id, color_rgb)

[0066] Vehicle(): Vehicle model factor configuration(); asset_id: Blueprint ID, such as vehicle.tesla.model3, specifies the vehicle's shape, size, and dynamic parameters; color_rgb: RGB body color, which can change the vehicle's appearance but not its dynamics, and is used to diversify training data.

[0067] Step S520: Dynamically control environmental factors through the time series interpolation engine. The time series interpolation engine parses all functional parameters at the start of the simulation task.

[0068] Step S521: Initialize a time series interpolator engine. At each time step of the simulation, based on the current simulation timestamp t and the preset function type (such as linear interpolation), accurately calculate the instantaneous values ​​(instantaneous target values) of all parameters at that moment. Use the linear interpolation function to realize the continuous change of parameters such as weather, illumination, and vehicle speed. Support multiple motion modes, including uniform speed, uniform acceleration, and variable acceleration.

[0069] Reference Figure 5 The diagram illustrates a function model of precipitation variation over a period of time. The four curves represent four different interpolation functions, illustrating how values ​​transition from initial to final values, thus producing different visual effects. Specifically, LINEAR represents uniform change, where the value increases at a constant rate over time; the curve is a straight, sloping line, suitable for stable scenarios without acceleration. EASE_IN represents slow initial change, then gradual acceleration, reaching its fastest point at the end; the curve is flat at the beginning, then gradually increases in slope, reaching its steepest point at the end; suitable for objects that need to overcome inertia to start, such as cars starting or projectiles being fired, giving the feeling of force exerted from a standstill. EASE_OUT represents rapid initial change, then gradual deceleration until a slow stop at the end; the curve is very steep at the beginning, then gradually decreases in slope, becoming flat at the end; suitable for simulating the effect of objects stopping naturally due to friction or resistance, such as braking or a thrown object landing, giving the feeling of natural deceleration and a smooth ending. S_CURVE (S-shaped curve) is a combination of EASE_IN and EASE_OUT. The process is: slow start → acceleration → constant speed → deceleration → slow end. The curve is characterized by a smooth "S" shape, which is closer to the motion of objects in the real world, making the motion look very natural, smooth and textured.

[0070] Reference Figures 1 to 4Step S522: Finally, these calculated instantaneous values ​​will be updated in real time to the scene's environment controller or the vehicle's PID controller via the simulation platform's API (such as the autonomous driving API), thereby achieving a visually smooth and continuous programmed gradual change effect, and thus simulating the gradual change process in the real physical world. Specifically, a cascaded PID controller or model predictive control (MPC) algorithm is used to achieve precise longitudinal and lateral control of the intelligent agent; a waypoint-based high-precision map path tracking algorithm is integrated to enable the intelligent agent to drive autonomously and compliantly in the road network.

[0071] The mechanism for continuous gradient technology: Taking the gradient mechanism of fog concentration as an example, firstly, the gradient function is defined as a configurable JSON template file, and so on. "fog_density":{"function_type": "LINEAR_INTERPOLATION", "start_time":0.0, "end_time": 10.0, "start_value": 0.0, "end_value": 0.8}

[0072] `function_type`: Indicates the type of function whose parameters change over time, such as linear, gradual ingress, gradual egress, S-shaped smoothing, etc. `start_time` / `end_time`: Indicates the time interval during which the parameter changes occur. `start_value` / `end_value`: Indicates the starting and ending values ​​of the function change, used to define the gradual range. `duration`: Indicates the total duration of the function change, providing a unified time reference for dynamic changes in speed or weather.

[0073] Step S530: Use a deterministic synchronization barrier mechanism to synchronously acquire data from the multi-view perception system. The deterministic synchronization barrier mechanism operates in a simulation synchronization mode forcibly initiated by the client. The client completely controls the evolution of simulation time by sending tick commands, ensuring zero-deviation alignment of multi-view image data at the acquisition time point. Specifically, after the client sends a tick command to drive the server to calculate a fixed time step, the client enters a blocking "synchronization barrier" waiting state. This barrier continuously listens for and collects data returned from all sensors at that time step until it confirms that it has received a complete set of multi-view perception system data from the same calculation before releasing it. Furthermore, the deterministic synchronization barrier mechanism also incorporates timeout protection logic to prevent permanent system blocking due to the loss of data from a single sensor, ensuring robustness for large-scale generation. This provides perfectly aligned input data for downstream algorithms such as multi-view data fusion and BEV (bird's-eye view) space construction, eliminating the potential for poor model training results due to asynchronous data at the source.

[0074] Step S600: Data Acquisition.

[0075] Step S610: During the vehicle's movement, five virtual cameras simultaneously acquire image data. The images are automatically converted to a standard format. After the acquired synchronous data packets are processed, they are automatically saved into the created structured hierarchical storage directory.

[0076] Multiple scene description files are collected from the multi-view perception system to generate a structured and traceable traffic scene dataset.

[0077] The above steps use multi-factor decoupling control and combination algorithms to independently describe key factors in traffic scenarios (such as map area, vehicle type, weather conditions, speed mode, spatial layout, etc.), and automatically instantiate scenarios through a factor combination engine to achieve rapid construction of large-scale scenarios.

[0078] Optionally, it also includes step S700: resource recycling.

[0079] Step S710: After each scene is completed, the system automatically destroys all simulation objects (vehicles, cameras) created in this task, completely releasing computing resources. It then retrieves a scene description file from the task queue, seamlessly switches between scenes, and repeats the execution, achieving unattended batch data generation.

[0080] Step S720: The data orchestration engine retrieves the next scene description file from the task queue and automatically and seamlessly repeats the entire process described above until the queue is empty, thus achieving large-scale, unattended GUI-controlled data generation.

[0081] The implementation principle of the multi-dimensional decoupled traffic scene data generation method for autonomous driving training in this application embodiment is as follows: Based on a digital twin simulation environment (autonomous driving), a modular automated data generation process is designed, supporting users to flexibly configure traffic environment parameters, including town / village maps, vehicle types (cars, SUVs, trucks, etc.), vehicle body color (customizable multiple schemes), driving speed (constant speed, acceleration, deceleration), and weather conditions (sunny, rainy, foggy, snowy, etc.). By constructing a multi-view synchronous acquisition architecture composed of surround-view cameras and top-view cameras, this invention can acquire high-resolution vehicle traffic images from multiple angles such as front view, rear view, left view, right view, and top view in real time in the simulation environment, supporting continuous video frame acquisition at 10-30 FPS, with configurable resolution (720P / 1080P, etc.), thereby achieving high-quality, customizable, and continuous generation of massive traffic image data.

[0082] This invention uses a parameterized environmental factor description method. For example, the weather factor is defined as Weather(precipitation, cloudiness, fog_density, sun_azimuth_angle); the vehicle speed factor is defined as Velocity(profile_type, start_velocity, end_velocity, duration); and the vehicle type factor is defined as Vehicle(asset_id, color_rgb).

[0083] Structured scene descriptions using JSON allow for flexible definition and support static configuration as scalar values ​​or dynamic gradual control as time functions. Users or upper-level programs can flexibly define the scenes to be generated by writing description files. When parameters are defined as functions, the time-series interpolation engine calculates the instantaneous value of the parameter based on the current simulation timestamp during simulation execution and updates it in real time through the simulation environment control interface. The JSON-based structured scene descriptions allow users to directly define and adjust environmental factors without modifying the underlying simulation code, thus achieving configurability, scalability, and reproducibility of the scene generation process. For example, a "rainy day + acceleration + black Model 3" scene can be described as follows:

[0084] {

[0085] "scene id": "rainy_acceleration test",

[0086] "factors":{

[0087] "weather":{"precipitation":0.9,...},

[0088] "velocity":{"profile_type":"LINEAR_ACCELERATION", "start_velocity":5.0, ...

[0090] }

[0091] User-demand-driven combinatorial scheduler: To achieve automated and systematic combination of factors, this invention introduces a pluggable combinatorial strategy engine. The scheduler first defines several discrete levels for each factor parameter. When the user seeks high-efficiency coverage, the scheduler calls an orthogonal experiment algorithm to optimize the combination of the discrete levels of the parameters, generating a scene description file with a minimal complete set. When the user performs deep boundary testing, the scheduler switches to full factor traversal mode or Monte Carlo random mode.

[0092] High-dimensional factor decoupling and orthogonal combination capabilities: This invention achieves factor-level decoupling control of four core dimensions—environment, behavior, space, and entities—in traffic scenarios. Through a scheduling engine, it can systematically and automatically generate single-factor variables and multi-factor combination scenarios. Effects: This enables external models to learn complex coupled scenarios and allows for more robust single-factor stress testing (e.g., testing the performance degradation curve of the perception algorithm by only changing fog concentration), providing crucial data support for model attribution analysis and credibility assessment.

[0093] Continuous temporal evolution of dynamic scenes: This invention breaks through the limitations of discretized environmental parameters and abrupt switching in traditional simulations. It adopts a functional sequence-driven approach to achieve continuous and smooth transitions of key factors such as weather, lighting, and vehicle speed. Results: The generated data better conforms to the gradual changes in the real physical world, effectively training and validating the model's temporal stability and adaptability during dynamic changes, greatly improving the realism and effectiveness of the simulation scene.

[0094] Deterministic Spatiotemporal Synchronization Mechanism: Through forced synchronization mode and synchronization barrier design, this invention establishes a deterministic multi-sensor data synchronization acquisition framework, ensuring zero-deviation alignment of multi-view image data at the acquisition time point. Effect: It provides perfectly aligned input data for downstream algorithms such as multi-view data fusion and BEV (bird's-eye view) space construction, eliminating the potential for poor model training performance due to asynchronous data at the source.

[0095] A fully automated production line with a closed-loop process: This invention integrates scene configuration, factor combination, dynamic execution, data acquisition, and structured storage into an automated closed-loop system that requires no manual intervention. Results: It significantly improves the production efficiency of large-scale, diverse, and structured datasets, achieving low-cost, high-efficiency data generation services.

[0096] Secondly, embodiments of the present invention also provide a multi-dimensional decoupled traffic scene device, including a memory storing a computer program and a processor. When the processor executes the computer program, it implements the multi-dimensional decoupled traffic scene data generation method for autonomous driving training described in the first aspect.

[0097] The above embodiments are merely illustrative of several implementation methods of this disclosure, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of this disclosure, and these modifications and improvements all fall within the protection scope of this disclosure.

Claims

1. A method for generating multi-dimensional decoupled traffic scene data for autonomous driving training, the method comprising: The method comprises the steps of: a user pre-defines a high-level configuration file, which is used to generate multiple scene description files based on a combination strategy engine, and the scene description files define multiple environmental factors in a structured manner, including at least one of a map, weather, illumination, time, vehicle speed, vehicle type, vehicle color, and traffic density; an automatic driving simulation platform service is started by initializing a digital twin simulation environment, a simulation environment connection is established, a high-precision geographic information model is loaded, and a global synchronization mode is enabled, the scene description files are loaded and parsed, and corresponding high-precision road and topological data are obtained; according to the configuration of the parsed scene description files, vehicle models of a specified type and color are loaded from a blueprint library of the simulation platform, and vehicles are instantiated, and vehicle positions are instantiated by a random point screening algorithm, and a multi-view perception system is deployed in a coordinate system of the instantiated vehicle body; the environmental factors are dynamically controlled by a time series interpolation engine to realize continuous and smooth evolution of the environmental factors; and data of the multi-view perception system is synchronized by using a deterministic synchronization barrier mechanism. The data of the multi-view perception system is collected, and the collected data of the multi-view perception system is used to generate a structured and traceable traffic scene dataset, which is automatically saved.

2. The multi-dimensional decoupled traffic scene data generation method for automatic driving training according to claim 1, characterized in that: The combination strategy engine generates multiple scene description files, which comprises the steps of: a plurality of discrete levels are defined for each environmental factor; a plurality of scene description files are automatically generated according to business requirements and manually defined / orthogonal tables, and each scene description file corresponds to a test case; the scene description files are batch scheduled to realize systematic coverage of the environmental factors.

3. The multi-dimensional decoupled traffic scene data generation method for automatic driving training according to claim 1, characterized in that: The time series interpolation engine comprises the steps of: in each simulation time step, an instantaneous value of a parameter is calculated according to a current simulation timestamp and a preset function type; the instantaneous value is updated to a scene environment controller or a vehicle controller in real time through an API of the simulation platform to realize continuous and gradual change of the environmental factors.

4. The multi-dimensional decoupled traffic scene data generation method for automatic driving training according to claim 1, characterized in that: The deterministic synchronization barrier mechanism comprises the steps of: after a fixed time step of simulation calculation, a client enters a blocked waiting state until a complete set of multi-view data from the same calculation is received, and then the client is released; timeout protection logic is built in to prevent permanent blocking of the system due to loss of data from a single sensor; multi-view image data is aligned without deviation at a collection point.

5. The multi-dimensional decoupled traffic scene data generation method for autonomous driving training according to claim 1, characterized in that: The multi-view perception system comprises virtual cameras deployed in front, rear, left, right, and top views of the vehicle body coordinate system, the virtual cameras are kept spatially consistent with the vehicle body through rigid body transformation, and multi-modal data can be synchronously collected.

6. The multi-dimensional decoupled traffic scene data generation method for autonomous driving training according to claim 1, characterized in that: The scene description file is defined in JSON format, supports static configuration or dynamic function parameter description, and realizes configurability, expandability, and reproducibility of the scene.

7. The multi-dimensional decoupled traffic scene data generation method for autonomous driving training according to claim 1, characterized in that: The method further comprises the steps of: a batch of scene description files are parsed and scheduled by an automated data arrangement and management engine, a hierarchical storage directory is generated according to a signature of the environmental factors, and automatic management and data tracing of cross-scene assets are realized.

8. The multi-dimensional decoupled traffic scene data generation method for autonomous driving training according to claim 1, characterized in that: The dynamic control of the environmental factors comprises the steps of: Linear interpolation function is used to realize the continuous change of parameters such as weather, light, and vehicle speed; Support multiple motion modes, including constant speed, uniform acceleration and variable acceleration; Apply parameter changes in real time through the simulation API to simulate the gradual change process of the real physical world.

9. The multi-dimensional decoupled traffic scene data generation method for autonomous driving training according to any one of claims 1-8, characterized in that: It also includes the steps of automatically destroying the simulation object after each scene is executed, releasing the computing resources, taking the next scene description file from the task queue, seamlessly switching and repeating the execution, realizing the unattended batch data generation, and completing the resource recycling.

10. A multi-dimensional decoupling traffic scenario device, characterized by: The application discloses a multi-dimensional decoupling traffic scene data generation method for automatic driving training, and comprises a memory storing a computer program and a processor. The processor executes the computer program to realize the multi-dimensional decoupling traffic scene data generation method for automatic driving training.

Images