An intelligent generation method for testing the autonomous capabilities of ground unmanned equipment

Abstract

The present invention discloses a method for intelligently generating test scenarios for the autonomous capabilities of ground unmanned equipment, comprising constructing a multi-level simulation scenario terrain model through noise-based height map generation, terrain erosion simulation, landform parameterized modeling and terrain fusion, and remote scene modeling; then modeling a highly realistic and complex natural environment based on the generation of dynamic lighting and weather conditions, terrain coverage primitives, unstructured roads, and static simulation elements of scene vegetation; and automatically generating test scenario cases for the autonomous reconnaissance and maneuverability capabilities of unmanned equipment by combining intelligent agent behavior models with dynamic triggering rules. The present invention intelligently constructs test scenarios for the autonomous reconnaissance and maneuverability capabilities of ground unmanned equipment in a controlled simulation environment, automatically constructing a variety of unmanned system simulation scenarios within seconds, and solving the current problems of low efficiency and high cost in constructing simulation scenarios for testing ground unmanned equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned equipment testing, and in particular to a method for intelligently generating a test scenario for the autonomous capability of ground unmanned equipment. Background Art

[0002] Comprehensive testing is a prerequisite for the large-scale deployment of ground unmanned systems. Among numerous evaluation methods, simulation testing allows for comprehensive, full-process training and testing of the autonomous capabilities of ground unmanned systems in a virtual environment. This helps enhance their autonomous reconnaissance and automated maneuvering capabilities, thereby improving their operational effectiveness. In the simulation process for ground unmanned systems, in addition to focusing on accumulating test mileage, it is also crucial to improve scenario quality. A diverse and comprehensive simulation scenario library has become the cornerstone for enhancing the intelligence level of ground unmanned systems.

[0003] However, existing technologies are limited in that they still rely primarily on manual CAD modeling for scene construction. Building large-scale outdoor scenes is time-consuming and labor-intensive, and they lack the ability to efficiently and autonomously generate diverse scenarios. Furthermore, high-quality simulation scenarios should offer broad coverage and be able to autonomously generate edge cases that are difficult to capture through real-world data collection. However, current simulation systems have yet to fully explore their capabilities in generating extreme scenarios. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology. To achieve the above purpose, an intelligent generation method for testing the autonomous capabilities of ground unmanned equipment is adopted to solve the problems raised in the above background technology.

[0005] A method for intelligently generating a test scenario for autonomous capabilities of ground unmanned equipment comprises the following steps:

[0006] S1. Construct a multi-level simulation scene terrain model through noise-based height map generation, terrain erosion simulation, landform parameter modeling and terrain fusion, and remote scene modeling;

[0007] S2, based on the generation of dynamic lighting and weather conditions, terrain cover primitives, unstructured roads, and static simulation elements of scene vegetation, high-realism and complex scene natural environment modeling;

[0008] S3. Combine the intelligent agent behavior model with dynamic triggering rules to automatically generate test scenario cases for the autonomous reconnaissance and maneuverability capabilities of unmanned equipment.

[0009] As a further solution of the present invention: the specific steps in S1 include:

[0010] S11. Perform multiple Perlin noise calculations using different frequency and amplitude parameters through multi-layer noise generation, and then superimpose the calculation results to generate a first-level terrain height map;

[0011] S12. Based on the obtained primary terrain height map, a water erosion algorithm with a progressive gradient is introduced to simulate the erosion of the terrain by water flow, thereby generating a secondary terrain height map.

[0012] S13, extracting height distribution from real-world landform data, introducing parameters such as mountain height, size, overall slope, surface attenuation, and edge attenuation to parameterize the landform, thereby obtaining a parameterized landform;

[0013] Then, the protrusion, saddle, and isolation parameters are introduced to fuse the generated secondary terrain height map with the parameterized landform to generate the tertiary terrain height map of the basic scene.

[0014] S14. Import the generated three-level terrain height map of the basic scene into the terrain module of UE5 to generate a core terrain model of the simulation scene. Taking the core terrain model as the center, expand and add the remote four-level scene that does not participate in the actual physical simulation to finally form the terrain model of the simulation scene.

[0015] As a further solution of the present invention: the specific steps in S11 include:

[0016] A layered approach is used, that is, multiple Perlin noise calculations are performed using different frequency and amplitude parameters, and then the results of multiple noises are combined;

[0017] The Perlin noise with fractal geometric characteristics generated by multiple superpositions is then used to simulate three-dimensional realistic terrain based on the limitations of terrain data on specific terrain.

[0018] As a further solution of the present invention: the specific steps in S12 include:

[0019] During the construction of the secondary terrain height map, multiple water erosion iterations are used to gradually modify the original terrain through erosion and sedimentation.

[0020] The key parameters of erosion rate, sedimentation rate, and evaporation rate are used to determine the path of water flow and changes in terrain.

[0021] As a further solution of the present invention: the specific steps in S2 include:

[0022] S21. Based on the graphics engine, build a variety of lighting and sky systems, as well as high-fidelity simulation of various weather conditions; and set the control parameters of each natural environment and classify them;

[0023] S22: For the type of field operation scene, based on the pre-built digital asset terrain material library, the corresponding scene terrain material is automatically assigned to the terrain on the simulation scene terrain model built in step S1;

[0024] For the generation of outdoor scene vegetation, the slope and altitude parameters of the simulated scene terrain constructed in step S1 are combined to dynamically adjust the weights to control the clustering and sparse distribution of vegetation. Multi-level vegetation ecological levels are set, and road networks are used as constraints for non-generable areas. The procedural content generation module of UE5 is used to automatically construct the scene vegetation distribution.

[0025] S23. For the generation of unstructured roads in the scene, based on the geometric structure of the simulated scene terrain model constructed in step S1, the scene road network is automatically constructed by fusing the randomization of path point parameters with the scene simulation terrain constraints.

[0026] As a further solution of the present invention: the specific steps in S21 include:

[0027] After setting each natural environment control parameter, classify it into random parameter group and grade parameter group;

[0028] The random parameter group is randomly controlled by a seed random generator; the level parameter group is designed to control the challenging level of the generated scenario based on the autonomous capability to be tested.

[0029] As a further aspect of the present invention: the random parameter set of illumination includes cloud type and cloud height, and sun type;

[0030] The lighting level parameter group includes the amount of cloud cover and the intensity of the sun;

[0031] The random parameter set of weather conditions includes rain, snow type, dust type, and wind direction and strength;

[0032] The weather condition level parameter group includes the intensity of rain, snow, fog and haze, and dust.

[0033] As a further solution of the present invention: the specific step of the challenge level is to normalize the level parameters of the lighting conditions and weather conditions and divide them into multiple challenge levels.

[0034] As a further solution of the present invention: the specific steps in S3 include:

[0035] S31. According to the constructed simulation scenario model, based on the mission type of the actual ground unmanned equipment, configure the unmanned equipment simulation scenario model;

[0036] S32: Autonomous reconnaissance capability test for ground unmanned equipment, constructing dynamic mission events for three scenarios: randomization of reconnaissance target location, brief exposure of reconnaissance target, and dynamic mixing of friendly and enemy forces;

[0037] Define the automatically generated terrain model and natural environment for each autonomous reconnaissance mission scenario, as well as the models and actions of the hostile target agents in each reconnaissance scenario;

[0038] S33, autonomous mobility capability testing for ground unmanned equipment, constructing dynamic mission events for three scenarios: unmanned vehicle logistics supply transportation, unmanned vehicle rapid mobility rescue, and satellite denial positioning information loss;

[0039] Define and automatically generate corresponding terrain models and natural environment styles for each autonomous maneuvering mission scenario, as well as static and dynamic obstacles and threat areas in the maneuvering scenario.

[0040] As a further solution of the present invention: the specific steps of constructing the reconnaissance target in S32 include:

[0041] The construction of reconnaissance targets in simulation scenarios includes multi-type target settings, target dynamic behavior and camouflage methods to simulate the complex conditions of a real battlefield;

[0042] The reconnaissance target has various movement modes, including stationary state, slow movement, and fast movement. The stationary target is usually located in a bunker or a hidden position in a building.

[0043] Compared with the prior art, the present invention has the following technical effects:

[0044] The above technical solution, through the intelligent generation method of autonomous capability test scenarios, gets rid of the reliance on manual construction of simulation scenarios, greatly reduces the construction cost of unmanned equipment simulation scenarios, and can achieve automatic generation of test scenarios in seconds, significantly improving the construction efficiency of simulation scenarios.

[0045] It uses autonomous control to generate various types of terrain features such as plains and mountains, various terrain distribution elements such as forests and grasslands, diverse weather environments such as sunny days, rain and snow, sandstorms, as well as various scene dynamic events for reconnaissance missions and maneuver missions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings:

[0047] Figure 1 A schematic diagram of the steps of the intelligent generation method of the embodiment disclosed in this application;

[0048] Figure 2A schematic diagram of the implementation process of automatically generating a terrain model for a simulation scene according to an embodiment disclosed in this application;

[0049] Figure 3 This is a schematic diagram of the scenario dynamic agent modeling process and typical test scenario types of the embodiment disclosed in this application. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] Please refer to Figure 1 In an embodiment of the present invention, a method for intelligently generating a test scenario for autonomous capabilities of ground unmanned equipment includes the following steps:

[0052] S1. Construct a multi-level simulation scene terrain model through noise-based height map generation, terrain erosion simulation, landform parameter modeling and terrain fusion, and remote scene modeling;

[0053] In this embodiment, Figure 2 The figure shows a schematic diagram of the implementation process of the automatic generation of the terrain model of the simulation scene; the terrain model of the simulation scene is automatically generated by a procedural terrain construction method that mixes multi-layer fractal noise and parametric landforms. The method includes the following steps: generating the first-level terrain based on the height map of Perlin noise; generating the second-level terrain with a high-fidelity terrain erosion model; parametric modeling of typical landforms and merging it with the second-level terrain to generate the third-level terrain; and merging it with the static mesh terrain model at the far end of the scene to generate the fourth-level final terrain;

[0054] The specific implementation steps include:

[0055] S11. Perform multiple Perlin noise calculations using different frequency and amplitude parameters through multi-layer noise generation, and then superimpose the calculation results to generate a first-level terrain height map;

[0056] Specifically, if random values are used to fill the height map when automatically generating terrain, the terrain generated in this way will only be random spikes, not terrain. In this embodiment, Perlin noise is used to generate basic terrain height data. Perlin noise is constructed by superimposing multiple noise functions with different wavelengths and amplitudes.

[0057] Although the basic terrain height map generated by Perlin noise simulates natural noise to a certain extent, it still does not fully reflect the irregularities of natural noise. This implementation example further adopts a layered approach, that is, using different frequency and amplitude parameters to perform multiple Perlin noise calculations, and then combining the results of multiple noises;

[0058] The Perlin noise generated by multiple superpositions has strong fractal geometric characteristics. By limiting the terrain data according to specific terrain, it is possible to simulate three-dimensional realistic terrain.

[0059] The automatic generation of the first-level terrain height profile is controlled by frequency and amplitude parameters, thereby simulating different types of terrain profiles. The noise wave of mountainous terrain is simulated by a noise wave with larger amplitude and smaller frequency, while the noise wave with larger frequency and smaller amplitude is used to simulate flat terrain.

[0060] The degree to which the noise function is superimposed is controlled by the parameter multiplication factor. During the automatic generation of the first-level terrain height map, the larger the multiplication factor, the richer and finer the simulated terrain details, but also the greater the computational load.

[0061] S12. Based on the obtained primary terrain height map, a water erosion algorithm with a progressive gradient is introduced to simulate the erosion of the terrain by water flow, thereby generating a secondary terrain height map.

[0062] Specifically, in order to further simulate the impact of erosion and deposition processes of natural factors such as water flow in the real world on terrain changes, a progressive gradient water erosion algorithm is further introduced in the implementation example to simulate the flow of water droplets on the terrain surface and its erosion of the terrain, thereby obtaining a more natural and diverse secondary terrain height map.

[0063] The secondary terrain heightmap is constructed through multiple iterations of water erosion. Through erosion and deposition, the original terrain is gradually modified to make it more detailed and realistic. Key parameters such as erosion rate, deposition rate, and evaporation rate determine the path of water flow and the changes in terrain.

[0064] The key parameters such as erosion rate, deposition rate, and evaporation rate are encapsulated in interfaces, and the intensity of terrain erosion simulation can be controlled and generated by random number seeds through the interface;

[0065] S13, extracting height distribution from real-world landform data, introducing parameters such as mountain height, size, overall slope, surface attenuation, and edge attenuation to parameterize the landform, thereby obtaining a parameterized landform;

[0066] Then, the protrusion, saddle, and isolation parameters are introduced to fuse the generated secondary terrain height map with the parameterized landform to generate the tertiary terrain height map of the basic scene.

[0067] Specifically, the procedural terrain model generation method based on multi-layer fractal noise can realize terrain model generation, but due to the complexity of the real outdoor world, the scene terrain model constructed by the secondary terrain height map has little difference in regional expression, and the overall structure of the macro terrain is still relatively similar, which cannot perfectly express the terrain characteristics of the real world.

[0068] Therefore, the present invention innovatively further integrates real terrain data information based on the constructed secondary terrain height map, and performs parameterized modeling on the real terrain data to generate more realistic, complex and diverse terrain features.

[0069] This implementation example uses mountain terrain as an example to demonstrate the process of parameterizing real terrain.

[0070] From the public real mountain terrain DEM database, the mountains with different morphologies are parameterized by linear interpolation to form various parameterized mountain templates;

[0071] Each type of mountain template is controlled by mountain height, size, overall slope, surface attenuation and edge attenuation parameters. During the automatic mountain generation process, when a mountain type is selected from the template library, the mountain's morphology and erosion degree are controlled by this parameter group;

[0072] The generated parameterized landform height map is merged with the secondary terrain height map. The protrusion, saddle, and isolation parameters are used to control the relative position of the parameterized landform in the secondary terrain map. During the height fusion process of the overlapping area, the highest value is used as the final value of the generated tertiary terrain height map.

[0073] S14, importing the generated three-level terrain height map of the basic scene into the terrain module of UE5 to generate a core terrain model of the simulation scene, and expanding and adding the remote four-level scene that does not participate in the actual physical simulation with the core terrain model as the center, to finally form the terrain model of the simulation scene;

[0074] Specifically, the generated terrain height map is imported into the simulation engine to generate the final scene terrain model. This implementation example takes the UE5 engine as an example, imports the three-level terrain height map into the engine's "landscape" mode, and divides and blocks it to generate the terrain model of the simulation scene.

[0075] The combat scenario of ground unmanned equipment is a typical large-scale outdoor scenario. Due to the limitation of computer memory, if the third-level terrain includes the extremely remote terrain that does not participate in the physical simulation interaction, it will cause the simulation operation efficiency to decrease.

[0076] To solve this problem, in this implementation example, the generated three-level terrain is used as the core simulation scene, and the preset terrain model is expanded outward. The preset terrain model exists in the form of a fixed static mesh. The core scene generated by the three-level terrain supports the automatic generation of subsequent scene elements and the simulation test tasks of ground unmanned equipment. The extended scene does not participate in the physical interaction of ground unmanned equipment and only exists as a visual rendering.

[0077] The terrain model can be expanded at multiple levels, with each level adjusting the LOD (Level of Detail) based on the specific core simulation scenario. Typically, 2-3 LOD levels can be set for the terrain. For example, the first level retains 70% of the model's mesh face count for mid-range, and the second level drops to 30% for long distances. This significantly reduces GPU load when terrain detail decreases at long distances.

[0078] For the materials of the static mesh of the extended terrain, it is necessary to reuse the material instances of the main terrain as much as possible, or create simplified versions of the materials, such as removing parallax mapping and reducing texture sampling to reduce shader complexity.

[0079] For extended terrain that does not require interaction, simply disable collision in the static mesh properties; and set an impassable fence wall at the edge of the core terrain scene to prevent the passage of unmanned ground equipment. Since drones may participate in simulation tasks, for example, the combat space of fixed-wing drones may cover the entire map and do not need to interact with the ground terrain, no fence is set to block the drones.

[0080] S2, based on the generation of dynamic lighting and weather conditions, terrain cover primitives, unstructured roads, and static simulation elements of scene vegetation, high-realism and complex scene natural environment modeling;

[0081] In this embodiment, after the ground unmanned equipment completes the generation of the simulated scene terrain model, a highly realistic scene natural environment is further constructed based on the scene terrain. In order to test the autonomous capabilities of the ground unmanned equipment, static scene simulation elements such as scene lighting, weather, and environmental primitives are automatically generated in the early scene.

[0082] The specific steps in S2 include:

[0083] S21, based on the graphics engine, builds a variety of lighting and sky systems, as well as high-fidelity simulations of various weather conditions;

[0084] And classify each natural environment control parameter into random parameter group and graded parameter group;

[0085] The random parameter group is randomly controlled by a seed random generator; the level parameter group is oriented to the challenging level of the autonomous capability control generation scenario to be tested;

[0086] Specifically, the generation of lighting and weather is as follows: The autonomous reconnaissance capability test of ground unmanned equipment requires the construction of diverse lighting and weather conditions, the generation of camera, radar and other sensor data in different environments, and the comprehensive robustness testing of algorithm models such as target recognition and target tracking.

[0087] This implementation example is based on the UE5 engine, a fifth-generation game engine developed by UE5. It designs the natural environment system for the autonomous capability test scenario of ground unmanned equipment, providing rich and diverse natural environment conditions for autonomous reconnaissance capability testing.

[0088] The generation of lighting first involves the realization of sky effects. This implementation example is based on a graphics rendering system to achieve the generation and control of natural elements such as clouds and sun luminous bodies.

[0089] This implementation example is designed using the BP_Sky_Sphere blueprint and Ultra Dynamic Sky plug-in integrated into UE5 to accelerate the development of the natural environment system.

[0090] The parameters of each sky effect are classified into random parameter groups and level parameter groups. The parameter type of the random parameter group has little impact on the autonomous reconnaissance capability of unmanned equipment, while the level parameter group will directly affect the challenge of the autonomous reconnaissance capability test scenario.

[0091] The random parameter group of lighting conditions includes: cloud type (static cloud, 3D volume, 2D dynamic cloud, etc.) and cloud height, sun type (sun altitude, sun size), etc.

[0092] The lighting condition level parameter group includes: cloud cover (the higher the value, the higher the challenge level), sun intensity (the lower the value, the higher the challenge level), etc.

[0093] In this implementation example, weather generation is designed using UE5's Ultra Dynamic Weather plug-in, which can greatly simplify the implementation of various weather environments. This implementation example extracts the relevant parameter interfaces, including rain, snow, fog and haze, and dust.

[0094] The parameters of each weather effect are classified into random parameter groups and level parameter groups. The parameter type of the random parameter group has less impact on the autonomous reconnaissance capability of unmanned equipment, while the level parameter group will directly affect the challenge of the autonomous reconnaissance capability test scenario.

[0095] The random parameter group of weather conditions includes: rain, snow type (particle size, material, etc.), dust type (particle size, material, etc.), wind direction and strength, etc.;

[0096] The weather conditions level parameter group includes: the intensity of rain, snow, fog and dust (the higher the value, the higher the challenging level);

[0097] To improve rendering efficiency, this implementation dynamically adjusts the particle complexity and number based on the distance between the camera and the particles. Distant particles are simplified to a smaller amount of geometry or lower-resolution materials to reduce GPU computation and maintain performance when rendering a large number of particles simultaneously.

[0098] The natural environment of the scene is automatically generated, and the control parameters of light and weather may cross-control the same natural conditions. This implementation example sets the priority set of the control parameters generated by weather to be higher than the control parameters generated by light.

[0099] The level parameters of lighting conditions and weather conditions are normalized and divided into multiple challenge levels. In the automatic generation process of autonomous reconnaissance capability test scenarios, in addition to directly controlling each parameter, the challenge level can also be directly input and the value is randomly selected within the corresponding level parameter space. In the random parameter group, each control parameter is randomly generated if it is not specified.

[0100] S22: For field operation scenarios, based on the pre-built digital asset terrain material library, automatically assign the corresponding scene terrain material to the terrain based on the simulation scene terrain model constructed in step S1;

[0101] For the generation of outdoor scene vegetation, the slope and altitude parameters of the simulated scene terrain constructed in step S1 are combined to dynamically adjust the weights to control the clustering and sparse distribution of vegetation. Multi-level vegetation ecological levels are set, and road networks are used as constraints for non-generable areas. The procedural content generation module of UE5 is used to automatically construct the scene vegetation distribution.

[0102] S23, for the generation of unstructured roads in the scene, based on the geometric structure of the simulated scene terrain model constructed in step S1, the scene road network is automatically constructed by fusing the randomization of path point parameters with the scene simulation terrain constraints;

[0103] Specifically, the automatic construction of scene static simulation elements is as follows: static simulation elements include terrain coverage primitives, unstructured roads, scene vegetation elements, etc.

[0104] (1) Generation of Terrain Cover Primitives. The first step in the generation of terrain cover primitives is to set the terrain material. This implementation example uses the preset terrain material to directly assign it based on the scene type. The preset scene terrain materials include four categories: forest, grassland, desert, and snow.

[0105] This example uses the terrain cover primitives for a forest scene as an example. The creation of the preset terrain material relies on the rich data assets of the gaming industry. Terrain textures, as well as various digital assets such as grass, moss, and rock, are downloaded from the Quixel Bridge library. This includes a rich collection of albedo, normal, and distance maps. Over 20 types of grass and rock assets are configured, with information such as size variation ranges, occurrence probability, and distribution coverage set to ensure random and diverse generation of the terrain cover primitives. Wind and collision trampling effects are also applied to the grass vegetation meshes to achieve a more realistic physical simulation.

[0106] (2) Generation of scene vegetation elements. First, a digital resource library containing diverse vegetation forms is constructed. Digital resources can be obtained from digital asset libraries such as Quixel Bridge, or independently constructed using software such as Speed Tree. Such digital resources ensure that their forms conform to natural growth characteristics and generate multiple morphological variants and models of different growth stages for the same vegetation type to enhance scene diversity.

[0107] Automated vegetation generation is developed based on UE5's PCG functionality. First, the scene's vegetation-free areas, such as roads, are defined. Vegetation is not generated in these areas. A random noise algorithm is used to generate the spatial distribution of vegetation density in the target area. Dynamic weighting is adjusted based on terrain slope and altitude parameters to control the distribution of clustered and sparse vegetation. When the scene span is sufficiently large, the scene is divided into multiple vegetation groups, each associated with a specific vegetation combination and distribution constraints. Further design of ecological dependency rules between vegetation groups is performed, such as defining symbiotic or exclusive relationships between specific plants, to ensure that the generated results conform to natural ecological laws.

[0108] Automated vegetation layout generation is performed within a virtual terrain environment. First, terrain feature data, including height, slope, and simulated humidity, is extracted as the basic input for vegetation distribution. A layered generation strategy is employed to categorize vegetation by altitude and ecological level, such as a grassy rock layer at the bottom, a shrub layer at the middle, and a forest layer at the top, to avoid visual conflicts and optimize computational efficiency. For each vegetation level, vegetation instances are generated in batches within the area that meets ecological constraints, using predefined distribution rules defined using PCG.

[0109] (3) Generation of unstructured roads. Wild scenes contain a large number of unstructured roads. This solution uses UE5's Spline Mesh component to generate road structures based on the road network structure, and presets various material maps of the Spline Mesh, such as dirt roads, muddy roads, snowy roads, etc., to adapt to a variety of scene types;

[0110] The present invention realizes the automatic construction of scenario road network by integrating parameter randomization configuration and terrain constraint analysis. The system supports two modes: manual setting and random generation: manual mode accepts user-specified road starting point, end point, width and curvature range;

[0111] In random mode, road parameters are automatically generated within a specified geographic area. During the road planning phase, the system projects control points onto a DEM elevation map to analyze terrain characteristics. First, areas with sudden elevation changes between adjacent grids are identified. If these exceed a preset threshold, they are marked as impassable. Simultaneously, the terrain slope field is calculated based on 3D coordinate differences, marking impassable areas where the slope exceeds the engineering limit.

[0112] Road network paths are generated using an improved A* algorithm that incorporates a terrain penalty mechanism in addition to the traditional path length penalty. The algorithm evaluates the terrain slope, elevation change, and path curvature of candidate path points, and generates a terrain fitness index through weighted overlay, dynamically adjusting the path search direction.

[0113] During the path extension process, slope compliance and curvature continuity are simultaneously checked to ensure that the generated path conforms to real-world characteristics. Finally, the path is smoothed using B-spline curves to obtain a single path trajectory.

[0114] In the passable area of the scene terrain, starting points are randomly scattered within a certain distance range to generate path trajectories, and finally a three-dimensional road network model with both random characteristics and authenticity is output.

[0115] The automated generation of the natural environment for simulation scenarios frees scenario model designers from having to worry about the implementation details of the underlying technology. This allows for rapid generation of diverse and massive test scenarios while significantly reducing the development costs of these scenarios.

[0116] S3. Combine the agent behavior model and dynamic trigger rules to automatically generate test scenarios for the autonomous reconnaissance and maneuverability capabilities of unmanned equipment. The specific steps include:

[0117] like Figure 3 As shown, the figure shows the scene dynamic agent modeling process and the schematic diagram of typical test scene types;

[0118] S31. According to the constructed simulation scenario model, based on the mission type of the actual ground unmanned equipment, configure the unmanned equipment simulation scenario model;

[0119] In this embodiment, the unmanned equipment simulation scenario model is configured as follows:

[0120] There are many types of tasks for ground unmanned equipment. This implementation example takes the construction of a ground unmanned vehicle intelligent body as an example. The vehicle motion simulation uses the vehicle module to configure the simulation equipment model, which includes the parameters of each subsystem assembly, the digital model and material mapping of the equipment, the animation blueprint of the simulation equipment, and the motion control mode of the unmanned equipment simulation model, as well as the path tracking mode, so that the unmanned equipment can automatically control the simulation motion according to the planned trajectory.

[0121] To ensure autonomous reconnaissance capabilities for unmanned equipment, the simulation equipment configuration should focus on modeling the simulated sensors for the equipment. This implementation example configures an optical sensor simulation model for unmanned equipment based on the UE5 SceneCapture2D camera component. First, based on the actual sensor position distribution of each device, the SceneCapture2D camera component is configured to the same position and adjusted to consistent parameters such as the field of view. A RenderTarget class is created in UE, and parameters such as the size and color mode of the generated image are configured. Each RenderTarget is assigned to the corresponding simulated camera. Finally, SceneCapture2D uses the RGB values obtained through the image rendering pipeline to synthesize the visible light camera image. The obtained depth D value will be used as the output image of the depth camera. For the infrared camera, the infrared Post Process material is assigned to the simulated camera, and different Post Process material effects are set for the detected target. The post-processed scene image presents an infrared effect and is output to the corresponding RenderTarget.

[0122] To ensure autonomous maneuverability of unmanned vehicles, simulation configuration should prioritize their motion performance. This implementation example builds motion control for the simulated equipment based on the ChaosVehicle plugin for UE5. First, based on the actual overall parameters of the actual equipment, the simulation model's parameters, such as vehicle mass, center of mass position, and height, are configured. Then, parameters such as the powertrain, transmission, brake assembly, steering assembly, and suspension assembly are configured for the simulation model's power-speed-torque curve, transmission mode and ratio, brake acceleration, wheel steering ratio, and suspension stiffness, ensuring the simulation model closely simulates the actual vehicle's motion as closely as possible.

[0123] S32: Autonomous reconnaissance capability test for ground unmanned equipment, constructing dynamic mission events for three scenarios: randomization of reconnaissance target location, brief exposure of reconnaissance target, and dynamic mixing of friendly and enemy forces;

[0124] Define the automatically generated terrain model and natural environment for each autonomous reconnaissance mission scenario, as well as the models and actions of the hostile target agents in each reconnaissance scenario;

[0125] In this embodiment, the basic environment is set up. The construction of reconnaissance targets in the simulation scenario includes multiple target types, target dynamic behavior, and camouflage methods, simulating the complex conditions of a real battlefield. Reconnaissance targets have various movement modes, including static, slow movement, and fast movement. Static targets are often located in concealed locations such as bunkers and buildings, increasing the difficulty of detection. Slow-moving targets simulate patrol teams or standby vehicles, while fast-moving targets test the recognition system's real-time response capabilities, especially in brief exposure situations, requiring the system to quickly identify and locate the target.

[0126] The scenario features a variety of target types, including enemy infantry, various vehicle types (such as armored vehicles, transport vehicles, and tactical command vehicles), and low-flying drones. Corresponding simulation models are pre-stored in the simulation model library and automatically invoked based on the type of dynamic event. The same type of vehicle is assigned a variety of paint schemes and camouflage patterns. For example, in the jungle scenario generation, the reconnaissance vehicle was painted military green to increase the challenge of the test scenario.

[0127] This implementation example takes three scenario tasks as examples: random reconnaissance target location, brief exposure of reconnaissance targets, and dynamic mixing of friendly and enemy forces. It demonstrates the automatic generation method of specific mission scenarios in the unmanned equipment autonomous reconnaissance capability test, which is used to evaluate the recognition effect and task completion efficiency of ground unmanned equipment in diverse scenarios.

[0128] (1) Randomization of reconnaissance target locations. Ground unmanned vehicles need to detect targets in an area under unknown target distribution conditions. This tests the area search capability of the unmanned vehicles, especially their recognition capabilities under conditions of limited field of view in complex natural environments.

[0129] In the test scenario, the initial positions of all enemy targets are randomly distributed in different terrain areas, and the distribution positions are regenerated for each test to simulate the unpredictability of target distribution in a real battlefield.

[0130] Automatically create plains and various types of mountain terrain based on step S1, and assign texture maps to the terrain;

[0131] In addition to the differences in topography and landforms, mountainous terrain further generates multiple types of scene vegetation distribution elements based on the automated construction of scene static simulation elements in step S22, forming forest scenes, shrub scenes, grassland scenes, barren mountain scenes, etc. On plain terrain, the road network generation automatically constructs scene roads and assigns road materials such as dirt roads. Preset forest vegetation is loaded on both sides of the scene road network based on the vegetation generation in step S22, and the distribution of scene static elements in other areas is randomly controlled and generated by the system.

[0132] Based on the generation of illumination and weather conditions in step S21, the scene conditions controlled by the random parameter group are controlled by a random number generator, and the scene conditions controlled by the level parameter group are randomly generated within the specified scene challenge level. The scenes with high challenge levels include

[0133] The parameters of each scene can also be adjusted manually;

[0134] The challenge level of the test scenario is controlled by the level parameter group in the natural environment generation process on the one hand, and by the location of the detected target on the other hand. The area where the detected target appears is set, and the probability of the target appearing in each area is controlled.

[0135] Low-level challenging target areas, such as areas without forest cover on plain terrain, use a random number controller to control the generation of reconnaissance targets; high-level challenging target areas include areas where targets are hidden by bushes, rocks, or partially obscured by trees.

[0136] Control the frequency of target appearance in the target appearance area with different probability densities, and randomly control the target position and number to generate reconnaissance targets;

[0137] The camouflage of the reconnaissance target is further set according to the material of the test scene. For high-challenge levels, such as dark green woodland camouflage for reconnaissance targets in grassland and forest scenes, desert spotted camouflage for reconnaissance targets in neighboring and desert scenes, and snow camouflage for reconnaissance targets in snowfield scenes, etc.

[0138] During the test, there were no restrictions on the movement trajectory of the unmanned equipment, which was allowed to move freely in the entire scene. The target detection rate and search time were used to test and evaluate the autonomous reconnaissance capability of the unmanned equipment.

[0139] (2) Brief exposure of reconnaissance targets. A specific target is exposed to the field of view for a short period of time at a random location. For example, the target appears briefly behind a cover such as a rock or tree and then quickly hides. This task tests the system's recognition and response capabilities within limited information and time.

[0140] Test evaluation indicators: recognition reaction time, short-term recognition accuracy and capture success rate during the exposure period.

[0141] Automatically create a scene terrain model based on step S1, and automatically assign terrain material based on step S22;

[0142] The unmanned ground equipment to be tested remains in a fixed position and maintains a fixed reconnaissance area;

[0143] The visible area of the ground unmanned equipment reconnaissance camera sensor is projected onto the test scene based on the sensor space transformation matrix to obtain the generation area of the target to be tested;

[0144] In grassland and low shrub scenes, the target will perform the following actions: lying down and then standing up; in forest scenes, the target will hide and run, like in a dense forest; in mountain scenes, the target will run and hide behind rocks or other cover. Based on the target generation rules, the following strategy is used to automatically generate targets:

[0145] In grassland and low shrub scenes, targets are randomly generated in the spawning area determined by the reconnaissance camera;

[0146] The target in the forest scene first obtains the location of vegetation based on the scene generated in S22, and uses a certain density value as a threshold to further divide the dense forest area. Within the generated area determined by the reconnaissance camera, the target randomly generates outside the divided forest area at a short distance and quickly enters the woods scene to hide;

[0147] The target in the mountain scene first obtains the position of the cover in the scene based on the scene generated in S22, and extracts the cover whose overall size is larger than the target size as the target's hiding cover. The target is finally generated to the side and rear of the reconnaissance camera position in the area determined by the camera; when the test task is executed, the target quickly hides behind the cover.

[0148] (3) Dynamic mixing of friendly and enemy forces. In the test scenario, friendly and enemy targets are mixed to simulate a complex combat environment, thereby testing the system's ability to distinguish between friend and foe and avoiding accidental injuries.

[0149] Test evaluation indicators: friend-enemy identification accuracy, recognition delay, and false recognition rate.

[0150] Automatically create a scene terrain model based on step S1, and automatically assign terrain material based on step S22;

[0151] The unmanned equipment to be tested remains in a fixed position and maintains a fixed reconnaissance area;

[0152] The visible area of the unmanned equipment reconnaissance camera sensor is projected onto the test scene based on the sensor space transformation matrix to obtain the generation area of the target to be tested;

[0153] Mix friendly and enemy targets with different camouflage patterns and paint schemes in the same scene area, control the frequency of friendly and enemy targets in the target generation area with different probability densities, and randomly control the target position and number to generate reconnaissance targets;

[0154] All targets are created by UE5's Mass AI into autonomous intelligent entities, and their movement trajectories have certain intersections and contacts, simulating the dynamic interweaving of friendly and enemy forces in the real world.

[0155] S33, autonomous mobility capability testing for ground unmanned equipment, constructing dynamic mission events for three scenarios: unmanned vehicle logistics supply transportation, unmanned vehicle rapid mobility rescue, and satellite denial positioning information loss;

[0156] Define and automatically generate corresponding terrain models and natural environment styles for each autonomous maneuvering mission scenario, as well as static and dynamic obstacles and threat areas in the maneuvering scenario.

[0157] In this embodiment, the basic environment setting and autonomous mobility capability test scenarios will focus on the construction of terrain conditions, covering various terrain types, terrain parameters, static and dynamic obstacles, impassable threat areas, and highly dynamic randomized emergencies.

[0158] Terrain types cover a variety of natural environments such as plains, mountains, forests and grasslands, making the test scenarios more realistic in simulation. Plains provide a wide space for movement, suitable for basic path planning and acceleration testing; mountains contain steep slopes and altitude differences, simulating the difficulty of climbing and descending in complex terrain; forest terrain has the characteristics of field of view obstruction and path interference, challenging the vehicle's obstacle avoidance ability in densely vegetated areas; grassland combines the openness of plains with the local complexity of forests to test the field of view expansion and recognition effects under different speed and distance conditions. The patent of this invention can directly control multiple parameters of the simulated terrain, including mountain height, altitude changes, vegetation cover, etc., making the scene more layered. For example, the mountain height and altitude changes create real undulating terrain through steep slopes and height differences, which increases the challenge to vehicle path planning and autonomous driving.

[0159] The automated maneuverability test scenario incorporates static obstacles such as abandoned military vehicles, sandbags, and chevaux de frise (barricades) to simulate common obstacles encountered in combat scenarios. Large obstacles like abandoned vehicles increase path obstruction, requiring the vehicle to identify and implement appropriate detour strategies. Sandbags and chevaux de frise not only block the path but also test the detour and obstacle avoidance capabilities of the vehicle's path planning system. Dynamic obstacle scenarios include vehicles suddenly changing lanes and randomly appearing pedestrians. These dynamic elements introduce uncertainty to the test, requiring the intelligent land warfare platform to quickly respond and make evasive decisions.

[0160] Threat zones include minefields and static or dynamic enemy radar zones. These impassable, high-risk areas require unmanned vehicles to identify and circumvent. Minefields are set as absolutely impassable, while radar zones simulate enemy reconnaissance threats based on dynamic or static conditions, increasing the risk avoidance requirements in path planning. Furthermore, the dynamic and unpredictable environment introduces scenarios such as sudden destruction of bridges or roads, necessitating rapid emergency response and rerouting for unmanned ground vehicles facing sudden path disruptions. These random changes in the environment significantly increase the complexity of the simulation scenarios, thereby comprehensively testing the intelligence and robustness of the system's autonomous maneuvering capabilities in unexpected situations.

[0161] This implementation example is based on environmental condition settings and takes three scenarios: logistics supply transportation, rapid mobile rescue, and satellite denial positioning information loss as examples. It demonstrates the automatic generation method of specific mission scenarios in the unmanned equipment autonomous mobility capability test, which is used to evaluate the autonomous mobility planning capability and task completion efficiency of ground unmanned equipment in various scenarios.

[0162] (1) Ground unmanned vehicle logistics supply transportation. Ground unmanned vehicles need to transport logistics supplies in diverse terrain environments and ensure that the supply transportation mission is completed efficiently within the set time.

[0163] Test evaluation indicators: planning time from task receipt to path generation, obstacle avoidance response time to dynamic obstacles, and success rate of successfully completing material transportation within the specified time.

[0164] Automatically create plains and various types of mountain terrain based on step S1, and automatically assign terrain materials;

[0165] Automatically generate scene weather conditions based on step S21, where the weather conditions are mainly good weather such as sunny days;

[0166] Based on step S22, various primitives in the scene are first automatically created, and then various vegetation elements in the scene are automatically generated. In this task mode, an unstructured road network is further generated, and a ground unmanned equipment driving path is automatically created based on the road network, and the starting point and end point of the unmanned equipment movement are set.

[0167] Static obstacles (such as abandoned vehicles and sandbags) are randomly distributed around the driving path, and the frequency of obstacles in the set area is controlled with different probability densities. The location and number of obstacles are also randomly controlled.

[0168] The generation of dynamic obstacles uses a state machine model to simulate the appearance of vehicles and pedestrians in the environment within a specific time, and sets conditional triggers to randomly generate dynamic obstacles to increase the challenge of autonomous mobile logistics supply transportation scenario testing.

[0169] (2) Rapid mobile rescue by ground unmanned vehicles. Ground unmanned vehicles need to reach the target location and complete the rescue in a short time, and their emergency response capabilities in complex threat environments must be assessed.

[0170] Test evaluation indicators: response time from receiving the task to arriving at the target, accurate avoidance rate of threat areas, and time to regenerate the path in the event of sudden obstacles.

[0171] Automatically create plains and various types of mountain terrain based on step S1, and automatically assign terrain materials;

[0172] Automatically generate scene weather conditions based on step S21, where the weather conditions are mainly good weather such as sunny days;

[0173] Based on step S22, various primitives in the scene are first automatically created, and then various vegetation elements in the scene are automatically generated. In this task mode, it can be determined randomly or manually whether to use the generated unstructured road network.

[0174] Use simulation to generate threat areas (such as minefields and enemy radar areas) in the scene, randomly generate the location and shape of the threat areas, and send them to the autonomous mobility module of the unmanned equipment under test; randomly distribute static obstacles (such as abandoned vehicles and sandbags) in the scene, control the frequency of obstacles in the set area with different probability densities, and randomly control the location and number of static obstacles;

[0175] The state machine model for generating dynamic events sets a conditional trigger when the ground unmanned equipment drives to a specific area, randomly generates an impassable area in front, simulates an enemy attack that makes the original impassable area impassable, and examines the unmanned equipment's autonomous maneuvering and replanning capabilities.

[0176] (3) Loss of satellite positioning information. Ground unmanned vehicles need to use real-time mapping and positioning algorithms to successfully locate themselves in the event of a sudden loss of satellite positioning signals, and evaluate their repositioning capabilities in the event of electromagnetic interference satellite positioning interruption.

[0177] Test evaluation indicators: the accuracy of real-time repositioning of ground unmanned equipment after signal loss.

[0178] Automatically create plains and various types of mountain terrain based on step S1, and automatically assign terrain materials;

[0179] Automatically generate scene weather conditions based on step S21, where the weather conditions are mainly good weather such as sunny days;

[0180] Based on step S22, various primitives in the scene are first automatically created, and then various vegetation elements in the scene are automatically generated. In this task mode, an unstructured road network is further generated, and a driving path of the ground unmanned equipment is automatically created based on the road network, and connected to the autonomous path tracking control module of the unmanned equipment to be tested;

[0181] During the path tracking process of a ground-based unmanned vehicle, the location and duration of positioning information loss are controlled with varying probability densities. During the period of information loss, the real-time mapping and positioning algorithm of the mobility being tested is connected to obtain the repositioning information of the unmanned vehicle. Signal loss in the autonomous mobility test scenario of this invention is achieved by disabling the GetActorTransform function in UE5.

[0182] Although the embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents, and all should be included within the scope of protection of the present invention.

Claims

1. A method for intelligently generating test scenarios for autonomous capabilities of ground unmanned equipment, characterized in that: The following steps are involved: S1. Construct a multi-layered simulation scene terrain model through noise-based height map generation, terrain erosion simulation, landform parameter modeling and terrain fusion, and remote scene modeling. The specific steps include: S11. Perform multiple Perlin noise calculations using different frequency and amplitude parameters through multi-layer noise generation, and then superimpose the calculation results to generate a first-level terrain height map; S12. Based on the obtained primary terrain height map, a water erosion algorithm with a progressive gradient is introduced to simulate the erosion of the terrain by water flow, thereby generating a secondary terrain height map. S13, extracting height distribution from real-world landform data, introducing parameters such as mountain height, size, overall slope, surface attenuation, and edge attenuation to parameterize the landform, thereby obtaining a parameterized landform; Then, the protrusion, saddle, and isolation parameters are introduced to fuse the generated secondary terrain height map with the parameterized landform to generate the tertiary terrain height map of the basic scene. S14, importing the generated three-level terrain height map of the basic scene into the terrain module of UE5 to generate a core terrain model of the simulation scene, and expanding and adding the remote four-level scene that does not participate in the actual physical simulation with the core terrain model as the center, to finally form the terrain model of the simulation scene; S2. Based on the generation of dynamic lighting and weather conditions, terrain cover primitives, unstructured roads, and static simulation elements of scene vegetation, a highly realistic and complex natural environment modeling is performed. The specific steps include: S21. Based on the graphics engine, build a variety of lighting and sky systems, as well as high-fidelity simulation of various weather conditions; and set the control parameters of each natural environment and classify them; S22: For the type of field operation scene, based on the pre-built digital asset terrain material library, the corresponding scene terrain material is automatically assigned to the terrain on the simulation scene terrain model built in step S1; For the generation of outdoor scene vegetation, the slope and altitude parameters of the simulated scene terrain constructed in step S1 are combined to dynamically adjust the weights to control the clustering and sparse distribution of vegetation. A multi-level vegetation ecological hierarchy is set, and the road network is used as a constraint for non-generable areas. The procedural content generation module of UE5 is used to automatically construct the scene vegetation distribution. S23, for the generation of unstructured roads in the scene, based on the geometric structure of the simulated scene terrain model constructed in step S1, the scene road network is automatically constructed by fusing the randomization of path point parameters with the scene simulation terrain constraints; S3. Combine the intelligent agent behavior model with dynamic triggering rules to automatically generate test scenario cases for the autonomous reconnaissance and maneuverability capabilities of unmanned equipment.

2. The method for intelligently generating a test scenario for autonomous capabilities of ground unmanned equipment according to claim 1, characterized in that: The specific steps in S11 include: A layered approach is used, that is, multiple Perlin noise calculations are performed using different frequency and amplitude parameters, and then the results of multiple noises are combined; The Perlin noise with fractal geometric characteristics generated by multiple superpositions is then used to simulate three-dimensional realistic terrain based on the limitations of terrain data on specific terrain.

3. The intelligent generation method for testing the autonomous capability of ground unmanned equipment according to claim 1, characterized in that: The specific steps in S12 include: During the construction of the secondary terrain height map, multiple water erosion iterations are used to gradually modify the original terrain through erosion and sedimentation. The key parameters of erosion rate, sedimentation rate, and evaporation rate are used to determine the path of water flow and changes in terrain.

4. The method for intelligently generating a test scenario for autonomous capabilities of ground unmanned equipment according to claim 1, characterized in that: The specific steps in S21 include: After setting each natural environment control parameter, classify it into random parameter group and grade parameter group; The random parameter group is randomly controlled by a seed random generator; the level parameter group is designed to control the challenging level of the generated scenario based on the autonomous capability to be tested.

5. The intelligent generation method for testing the autonomous capability of ground unmanned equipment according to claim 4, characterized in that: The random parameter set for lighting includes cloud type and cloud height, as well as sun type; The lighting level parameter group includes the amount of cloud cover and the intensity of the sun; The random parameter set of weather conditions includes rain, snow type, dust type, and wind direction and strength; The weather condition level parameter group includes the intensity of rain, snow, fog and haze, and dust.

6. The method for intelligently generating a test scenario for autonomous capabilities of ground unmanned equipment according to claim 5, characterized in that: The specific steps of the challenge level are: normalizing the level parameters of the lighting conditions and weather conditions and dividing them into multiple challenge levels.

7. The method for intelligently generating a test scenario for autonomous capabilities of ground unmanned equipment according to claim 1, characterized in that: The specific steps in S3 include: S31. According to the constructed simulation scenario model, based on the mission type of the actual ground unmanned equipment, configure the unmanned equipment simulation scenario model; S32: Autonomous reconnaissance capability test for ground unmanned equipment, constructing dynamic mission events for three scenarios: randomization of reconnaissance target location, brief exposure of reconnaissance target, and dynamic mixing of friendly and enemy forces; Define the automatically generated terrain model and natural environment for each autonomous reconnaissance mission scenario, as well as the models and actions of the hostile target agents in each reconnaissance scenario; S33, autonomous mobility capability testing for ground unmanned equipment, constructing dynamic mission events for three scenarios: unmanned vehicle logistics supply transportation, unmanned vehicle rapid mobility rescue, and satellite denial positioning information loss; Define and automatically generate corresponding terrain models and natural environment styles for each autonomous maneuvering mission scenario, as well as static and dynamic obstacles and threat areas in the maneuvering scenario.

8. The method for intelligently generating a test scenario for autonomous capabilities of ground unmanned equipment according to claim 7, characterized in that: The specific steps of constructing the reconnaissance target in S32 include: The construction of reconnaissance targets in simulation scenarios includes multi-type target settings, target dynamic behavior and camouflage methods to simulate the complex conditions of a real battlefield; The reconnaissance target has various movement modes, including three states: stationary state, slow movement, and fast movement.

Images