Virtual simulation test closed-loop method and system for unmanned equipment
Through the closed-loop method of virtual simulation testing, a three-dimensional grid model was generated and the perspective was completed using NeRF technology to solve the problem of scene authenticity and scale conflict in unmanned equipment simulation testing, and efficient and comprehensive testing results were achieved.
Patent Information
- Application Number
- CN202510620708.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-14
- Publication Date
- 2025-08-15
AI Technical Summary
Traditional unmanned equipment simulation testing methods cannot fully cover complex environments, and the test results are out of touch with the real scene.
The virtual simulation test closed-loop method is adopted to collect actual test data, generate a three-dimensional grid model, combine high-precision maps, input obstacle information, and use NeRF technology to complete the perspective to form a closed-loop feedback mechanism for testing.
It realizes efficient and comprehensive testing of unmanned equipment in a virtual environment, improves the accuracy and comprehensiveness of the test, supports dynamic iterative optimization, and has good scalability.
Smart Images
Figure CN120493546A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of unmanned equipment simulation testing, and in particular to a closed-loop method and system for virtual simulation testing of unmanned equipment. Background Art
[0002] In recent years, with the rapid development of artificial intelligence, the Internet of Things, and autonomous driving technologies, unmanned equipment such as drones, autonomous vehicles, and robot dogs has become increasingly widespread in military and civilian applications. While these unmanned systems are becoming increasingly intelligent in key technologies such as perception, decision-making, and execution, their performance testing and evaluation in complex and changing environments still face numerous challenges, such as ensuring accurate environmental perception, reliable path planning, and real-time obstacle avoidance control. Therefore, efficient and comprehensive testing and evaluation of unmanned equipment is crucial.
[0003] Traditional simulation testing methods for unmanned equipment primarily include two approaches: 1. Closed-loop testing based on scenario description languages: Engineers manually construct test scenarios, but due to limitations in the number and diversity of scenarios, they struggle to cover the complexities of real-world environments, resulting in incomplete test results. 2. Playback of actual field test data: While capable of capturing real data, the lack of dynamic interaction between algorithms and data prevents a true reflection of the unmanned equipment's actual operational performance, leading to a disconnect between test results and real-world scenarios.
[0004] Therefore, a testing method that can achieve a closed-loop scenario is needed for simulation testing to resolve the conflict between the authenticity of the simulation scenario and the scale of the scenario. Summary of the Invention
[0005] The purpose of the present invention is to provide a closed-loop method and system for virtual simulation testing of unmanned equipment, which solves the problem that the test results of traditional unmanned equipment simulation testing methods are not comprehensive and are out of touch with real scenarios.
[0006] To achieve the above object, the present invention provides a closed-loop method for virtual simulation testing of unmanned equipment, comprising the following steps:
[0007] Collect actual test data and classify and save scene data segments based on action elements;
[0008] Use sensors to obtain depth and texture information of the scene, generate point cloud data and build a 3D mesh model, and combine it with high-precision maps to complete 3D static scene reconstruction;
[0009] Input static obstacle coordinate information and dynamic obstacle trajectory information into the 3D static scene and match it with existing map data;
[0010] The unmanned equipment is driven into the simulation initialization state using actual data. The equipment is tested using its algorithm. The perception module processes the simulation scene data and transmits it to the decision-making and control module, while also updating the equipment's position, posture, and motion status information.
[0011] After receiving the equipment status information, the simulation software backend module will synchronously update the 3D simulation scene and use NeRF technology to complete or synthesize new perspectives;
[0012] The obstacle information perceived by the virtual sensor at the next moment is transmitted back to the perception module, and the test is repeated until the end.
[0013] Among them, actual test data is collected, and scene data fragments are classified and saved based on action elements.
[0014] Action elements include six levels of classification: overall task layer, scene description layer, ground element layer, environmental climate layer, dynamic interaction layer and data interaction layer.
[0015] The sensor is used to obtain the depth and texture information of the scene, generate point cloud data and build a 3D mesh model, and then the 3D static scene is reconstructed in combination with a high-precision map. The specific steps include:
[0016] The texture information of the scene is obtained through the camera, and the depth information of the scene is obtained through the lidar;
[0017] Utilize target recognition algorithms to extract dynamic target information from acquired data and generate point cloud data of dynamic targets;
[0018] Perform pre-processing, surface reconstruction, mesh optimization and post-processing on point cloud data;
[0019] The same space is scanned repeatedly to complete the scene construction of the three-dimensional static space.
[0020] Among them, the point cloud data is pre-processed, surface reconstructed, mesh optimized and post-processed. The specific steps include:
[0021] Statistical filtering is used to remove noise points, multi-view point cloud registration is performed based on feature matching or iterative closest point algorithm, and voxel filtering is used to streamline point cloud data to improve processing efficiency.
[0022] Use Delaunay triangulation to construct the initial triangular mesh, fit the smooth surface and extract the triangular mesh through the moving least squares method, or use Poisson reconstruction to solve the Poisson equation to generate an implicit surface and then convert it into a triangular mesh;
[0023] Use the Laplace smoothing algorithm to adjust vertex positions to smooth the surface, and use the edge contraction algorithm to simplify the mesh and repair holes and topological errors;
[0024] Map photogrammetric or hand-painted textures onto mesh models, add lighting and material effects using graphics rendering libraries, and generate realistic 3D mesh models.
[0025] Among them, the static obstacle coordinate information and dynamic obstacle trajectory information are input into the 3D static scene and matched with the existing map data.
[0026] Obtain the precise location and attribute information of roads, buildings, and traffic signs based on high-precision map data, lidar point cloud data, and camera image data collected by sensors;
[0027] By matching the vehicle's position with a high-precision map, the coordinate information of static obstacles around the vehicle is determined;
[0028] The static obstacle coordinate information is input into the 3D static scene and matched with the existing 3D static map.
[0029] Among them, the static obstacle coordinate information and dynamic obstacle trajectory information are input into the 3D static scene and matched with the existing map data.
[0030] Process sensor data in real time and build a multi-hypothesis tracking model for dynamic obstacle tracking, including:
[0031] At time step t, the sensor observation set Z t ={z1, z2, ..., z Nt} and the target hypothesis set H t-1 Make an association;
[0032] Calculate the Mahalanobis distance d(z,h)=[(zz^) T S -1 (zz^)] 1 / 2 Evaluate the match between observations and hypotheses and construct a data association matrix, where z^ is the predicted observation of the target hypothesis h and S is the covariance matrix;
[0033] Generate target-observation association hypothesis combinations based on branch-and-bound algorithm;
[0034] Using the state transfer equation x t =F t x t -1+v t and observation model z t =H t x t +w t Perform status updates and predictions, where F t is the state transfer matrix, v t is the process noise, H t is the observation matrix, w t is the observation noise;
[0035] According to Bayes' rule, the posterior probability of each hypothesis P(h|Z 1:t )∝P(Z t |h)P(h|Z 1:t-1 );
[0036] Retain the high confidence hypothesis set H through pruning strategy t As a result of tracking;
[0037] Predict the trajectory, velocity and acceleration information of dynamic obstacles;
[0038] Input dynamic obstacle trajectory information into the 3D static scene.
[0039] Among them, after the simulation software backend module receives the equipment status information, it synchronously updates the 3D simulation scene and uses NeRF technology to complete or synthesize new perspectives.
[0040] The simulation software backend module is the basic platform for simulation testing, which is used to integrate the algorithm under test, three-dimensional static space, equipment status information and static environment map.
[0041] A virtual simulation test closed-loop system for unmanned equipment operates using the virtual simulation test closed-loop method for unmanned equipment.
[0042] The present invention provides a closed-loop method and system for virtual simulation testing of unmanned equipment. The method first extracts actual test data, performs action extraction based on action elements (such as static environment, dynamic targets, and equipment behavior), and classifies and stores scene data fragments. Next, sensors are used to obtain the scene's depth and texture information, generate point cloud data, and construct a three-dimensional mesh model. Combining high-precision maps and sensor data, the scene is reconstructed in a static three-dimensional space, and static obstacle coordinates and dynamic obstacle trajectory information are input. Subsequently, the unmanned equipment is driven into a simulation initialization state using actual data, and the equipment is tested using the algorithm of the equipment under test. The perception module processes the simulation scene data and transmits it to the decision-making and control module, while also updating the equipment's position, posture, and motion status information. After receiving the equipment status information, the simulation software backend module synchronously updates the three-dimensional simulation scene and uses NeRF technology to complete or synthesize new perspectives. Finally, the obstacle information data perceived by the virtual sensor at the next moment is transmitted back to the perception module, and the equipment's algorithm under test is tested until the test is completed. By combining virtual simulation technology with closed-loop testing principles, efficient and comprehensive testing of unmanned equipment in a virtual environment is achieved. This method collects and analyzes the unmanned equipment's operational data and control instructions in real time within a virtual simulation scenario, forming a closed-loop feedback mechanism that improves testing accuracy and comprehensiveness. Based on the test results, the unmanned equipment's control strategy is adjusted and optimized, enabling dynamic iterative testing and improving testing efficiency. The closed-loop testing method is also highly scalable, allowing for the addition of new test scenarios and test metrics based on actual needs to meet the testing requirements of unmanned equipment in diverse application scenarios. This approach resolves the conflict between simulation scenario realism and scale. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art.
[0044] Figure 1 It is an architectural block diagram of the simulation software of the present invention.
[0045] Figure 2 It is a simulation closed-loop data processing flow chart of the present invention.
[0046] Figure 3 It is a flowchart of the steps of the closed-loop method for virtual simulation testing of unmanned equipment of the present invention. DETAILED DESCRIPTION
[0047] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, but should not be understood as limiting the present invention.
[0048] See also Figures 1 to 3 ,in, Figure 1 It is an architectural block diagram of the simulation software of the present invention. Figure 2 It is a simulation closed-loop data processing flow chart of the present invention. Figure 3 It is a flowchart of the steps of the closed-loop method for virtual simulation testing of unmanned equipment of the present invention.
[0049] The present invention provides a closed-loop method for virtual simulation testing of unmanned equipment, comprising the following steps:
[0050] S101: Collect actual test data and classify and save scene data segments based on action elements;
[0051] Specifically, the action-based factor extraction process is as follows: Real-world LiDAR and cameras are used to collect information about the scene in which the equipment is operating, and specific behaviors are classified. The classification method draws on the scene classification method used in autonomous driving. The scene environment is classified based on task requirements, the speed of change of various factors in the scene and their priority influencing the task. It is divided into six layers. The first layer is the overall task layer (L1). This layer describes the tasks that the unmanned equipment can complete in the scene, such as intelligence reconnaissance, firepower strike, surveillance, and protection. This layer directly specifies the constraints imposed on the unmanned equipment by the scene. Since the tasks are fixed and rarely change, it is classified as the first level of the classification. The second layer is the scene description layer (L2). This layer describes the surface type and terrain complexity of the unmanned equipment's scene. It also includes the boundary size of the current environment, the number of special areas, and the degree of road structure. This layer describes the basic environmental conditions of the scene and provides a rough overview of the scene environment in which the unmanned equipment is located. Due to the wider application scope of unmanned equipment, the scene classification requires a larger scope than the autonomous driving classification system. Once the environment is determined, the above elements remain largely unchanged. However, since the environment is required to support the mission during testing, they are placed in the second layer. The third layer is the ground element layer (L3). This layer covers all static elements in the environment that interact statically or passively with the unmanned vehicle, including their status, such as ground material, type of traversable area, proportion of traversable area, slope of traversable area, changes in the road or route in the forward direction, type of special area, and height or depth of obstacles. This layer describes elements that are easier to describe when the unmanned vehicle moves within the scene. The routes described in the above classification refer to informal paths calculated by built-in programs that allow the unmanned vehicle to traverse, such as routes through grass or sand. However, unmanned vehicles are not subject to traditional traffic regulations during their missions, so traffic elements only account for a portion of the elements. The specific connotations of the same names can also vary. For example, special areas here refer to areas of the ground that affect movement or are significantly different from the larger environment, such as bumps, depressions, or excessively smooth surfaces. After the environment is determined, the impact of these elements on the unmanned vehicle will change depending on the progress of the mission, so they are placed in the third layer. The fourth layer is the environmental climate layer (L4), which describes all weather conditions relevant to the unmanned equipment mission, including heat levels, humidity, and optical visibility. Different weather conditions can cause slippery surfaces and blurred vision, which can affect the performance and effectiveness of unmanned equipment and, in turn, influence various decision-making processes. These factors affect unmanned equipment throughout the combat mission and change automatically over time, albeit slowly, hence their placement in the fourth layer.The fifth layer is the dynamic interaction layer (L5). This layer covers all dynamic elements or states in the environment that interact dynamically or actively with the unmanned equipment, such as the type and number of other vehicles, the type and number of people in motion, the type and number of animals in motion, the direction of motion relative to the vehicle, the states of static elements with abnormal motion, and electromagnetic signal interference, thermal radiation interference, nuclear radiation interference, and chemical gas corrosion interference that can harm the unmanned equipment. This layer describes the more difficult-to-describe elements in the unmanned equipment scenario. Due to the wide range of applications of unmanned equipment and the complex battlefield environment, a variety of moving elements or motion states are prone to occur. These elements and states may interfere with the movement of the unmanned equipment or even actively cause damage to the unmanned equipment. Due to the high degree of variability, these elements are placed in the fifth layer. The sixth layer is the data interaction layer (L6). This layer describes the means of providing information support to the unmanned equipment in the scenario. This layer describes all friendly data exchange, communication, and cooperation related to the unmanned equipment's combat mission, including various communication information and map data. Since unmanned equipment data interacts frequently and changes can even reach milliseconds, it is the most frequently changing element, so the above elements are placed in the sixth layer. This patent only processes the above-mentioned weather environment, obstacle status and the equipment's own status. Among them, static environment obtains weather and other information through data collection network information, dynamic targets are identified through yolo (camera) and laser point cloud (pointpillar), and the equipment's own behavior can be collected by collecting its own data. It is mainly based on the characteristics of the above-mentioned specific behaviors, which are automatically extracted through scripts and the actual test data is divided into data segments.
[0052] S102: Using sensors to obtain depth and texture information of the scene, generating point cloud data and constructing a 3D mesh model, and combining it with a high-precision map to complete 3D static scene reconstruction;
[0053] Specifically, the depth and texture information of the scene can be obtained through cameras and lidars, and the dynamic target information can be extracted through target recognition algorithms. The point cloud data of the scene is generated by dynamic targets manually or through professional software, and point cloud preprocessing is performed. Noise points are removed by using methods such as statistical filtering, and multi-view point cloud registration is performed by methods such as feature-based or iterative closest point (ICP). Voxel filtering is used to streamline data to improve processing efficiency. Surface reconstruction is then carried out, and the initial triangular mesh is constructed using Delaunay triangulation. The triangular mesh is extracted after fitting a smooth surface using the moving least squares method, or the implicit surface is obtained by solving the Poisson equation using Poisson reconstruction and then converted into a triangular mesh. Mesh optimization is then performed, and the vertex positions are adjusted to make the surface smoother with the help of the Laplace smoothing algorithm. The mesh is simplified using algorithms such as edge contraction, and topological errors such as holes are checked and repaired. Finally, post-processing is performed to map the texture obtained by photogrammetry or manually drawn onto the model, and graphics rendering libraries such as OpenGL are used to add lighting, material and other effects for model rendering, thereby generating a realistic three-dimensional mesh model. By completing the reconstruction of the scene in the static three-dimensional space and repeatedly scanning the same space, the scene construction of a limited range of the three-dimensional static space can be completed. The specific range is determined by the sensor's perception and detection range. After the construction is completed, the updated three-dimensional static space is stored.
[0054] S103: Inputting the static obstacle coordinate information and dynamic obstacle trajectory information into the three-dimensional static scene and matching it with the existing map data;
[0055] Specifically, the acquisition of coordinate information of static obstacles mainly depends on real high-precision maps and sensor data, among which the precise map information is acquired by surveying and mapping, and the static obstacles in the high-precision map are mainly imported after unifying the coordinate system. According to the high-precision map data, the precise position and attribute information of static obstacles such as roads, buildings, and traffic signs can be obtained. By matching the position of the vehicle with the high-precision map, the coordinate information of static obstacles around the vehicle can be obtained. By processing and analyzing the lidar point cloud data and the surrounding image information captured by the camera, the position and shape of the static obstacles can be identified, and their coordinate information can be obtained. The acquisition of dynamic obstacle trajectory information mainly depends on the real-time processing and analysis of sensor data. The acquisition of motion information such as the position and speed of dynamic obstacles is similar to the above-mentioned static scene. On this basis, a multi-hypothesis tracking model is established. First, at time step t, the sensor observation 2 is converted into 2 Set Z t ={z1, z2, ..., z Nt} and the existing target hypothesis set H t-1 To perform the association, we calculate the Mahalanobis distance d(z,h)=[(zz^) T S -1 (zz^)]1 / 2 (where z^ is the predicted observation of the target hypothesis h, and S is the covariance matrix) evaluates the match between the observation and the hypothesis and constructs the data association matrix. Then, based on algorithms such as branch and bound, possible hypothesis combinations are generated, each of which represents a target-observation association hypothesis. Next, the state transition equation x is used to t =F t x t -1+v t (F t is the state transfer matrix, v t is process noise) and the observation model z t =H t x t +w t (H t is the observation matrix, w t is the observation noise) to update and predict the hypothesis, and calculate the posterior probability P(h|Z) of each hypothesis according to Bayes' rule 1:t )∝P(Z t ∣h)P(h∣Z 1:t-1 Finally, the low-probability hypothesis is deleted through the pruning strategy, and the high-confidence hypothesis set H is retained. t This is used as the target tracking result for the current moment and passed to the next iteration. Dynamic obstacles are continuously tracked, their future motion trajectories are predicted, and their possible velocity, acceleration, and other information are calculated. After obtaining the static obstacle coordinates and dynamic obstacle trajectory information, these are input into the previously updated 3D static space, where the corresponding obstacles are generated.
[0056] S104: Using actual data, the unmanned equipment is driven into a simulation initialization state. The equipment is tested using its algorithm. The perception module processes the simulation scene data and transmits it to the decision-making and control module. The equipment's position, posture, and motion status information are also updated.
[0057] Specifically, actual data is used to drive the unmanned equipment into the simulation initialization state. After entering the state, the current frame data is used as the starting data, and the algorithm of the equipment under test is connected for testing. Among them, the algorithm under test mainly involves the perception, positioning, and control algorithms of the equipment. The perception module of the algorithm under test processes the simulation scene data after receiving it, that is, it processes the image and point cloud data, generates a target list and a passable area, and transmits it to the planning and control module. The perception module transmits the processed structured information to the decision-making and control module of the equipment, and transmits the trajectory points and motion status information of the vehicle at the next moment of the control module to the three-dimensional scene, completing the update of the equipment's position, posture and motion status information.
[0058] S105: After receiving the equipment status information, the simulation software backend module synchronously updates the 3D simulation scene and uses NeRF technology to complete or synthesize a new perspective;
[0059] Specifically, the simulation backend module is a basic simulation testing platform, primarily used to integrate the algorithm under test, a three-dimensional static space, equipment status information, and a static environment map. After the simulation software backend module receives the equipment status information, it also synchronously updates information about the equipment model's surrounding environment, obstacles, and other aspects of the 3D simulation scene. The equipment algorithm under test uses sensors such as the virtual camera and video camera onboard the equipment model to obtain information about dynamic and static objects within its perception range at the next moment. For obstacles and viewpoints that cannot be captured, NeRF complements or synthesizes new viewpoints. The synthesis process is as follows: NeRF first uses a deep neural network (MLP) to learn a continuous representation of the 3D scene. This representation includes the color and volume density of each spatial point. Then, for the viewpoints to be complemented or synthesized, NeRF samples 3D points along camera rays and inputs these points and their corresponding viewpoints into the neural network to obtain the color and density of each point. Finally, volume rendering technology is used to accumulate these colors and densities into a 2D image, resulting in an image with the new viewpoint.
[0060] S106: Transmit the obstacle information sensed by the virtual sensor at the next moment back to the perception module, and repeat the test until the end.
[0061] Specifically, the obstacle information data in the simulation scene perceived by the virtual sensor at the next moment is transmitted to the perception module, and the perception regulation and control modules in the equipment's tested algorithm are tested. This process is repeated until the final test is completed.
[0062] By combining virtual simulation technology with closed-loop testing principles, efficient and comprehensive testing of unmanned equipment in a virtual environment is achieved. This method collects and analyzes the unmanned equipment's operational data and control instructions in real time within the virtual simulation scenario, forming a closed-loop feedback mechanism that improves the accuracy and comprehensiveness of testing. Based on the test results, the unmanned equipment's control strategy is adjusted and optimized, enabling dynamic iterative testing and improving testing efficiency. Furthermore, the closed-loop testing method is highly scalable, allowing for the addition of new test scenarios and test metrics based on actual needs to meet the testing requirements of unmanned equipment in diverse application scenarios. This resolves the conflict between simulation scenario realism and scale.
[0063] The above disclosure is merely one or more preferred embodiments of the present application and is not intended to limit the scope of the present application. A person skilled in the art will understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present application are still within the scope of the present application.
Claims
1. A closed-loop virtual simulation test method for unmanned equipment, characterized in that: The following steps are involved: Collect actual test data and classify and save scene data segments based on action elements; Use sensors to obtain depth and texture information of the scene, generate point cloud data and build a 3D mesh model, and combine it with high-precision maps to complete 3D static scene reconstruction; Input static obstacle coordinate information and dynamic obstacle trajectory information into the 3D static scene and match it with existing map data; The unmanned equipment is driven into the simulation initialization state using actual data. The equipment is tested using its algorithm. The perception module processes the simulation scene data and transmits it to the decision-making and control module, while also updating the equipment's position, posture, and motion status information. After receiving the equipment status information, the simulation software backend module will synchronously update the 3D simulation scene and use NeRF technology to complete or synthesize new perspectives; The obstacle information perceived by the virtual sensor at the next moment is transmitted back to the perception module, and the test is repeated until the end.
2. The closed-loop virtual simulation test method for unmanned equipment according to claim 1, characterized in that: Collect actual test data and classify and save scene data segments based on action elements. Action elements include six levels of classification: overall task layer, scene description layer, ground element layer, environmental climate layer, dynamic interaction layer and data interaction layer.
3. The closed-loop virtual simulation test method for unmanned equipment according to claim 1, characterized in that: The sensor is used to obtain the depth and texture information of the scene, generate point cloud data and build a 3D mesh model. Combined with the high-precision map, the 3D static scene is reconstructed. The specific steps include: The texture information of the scene is obtained through the camera, and the depth information of the scene is obtained through the lidar; Utilize target recognition algorithms to extract dynamic target information from acquired data and generate point cloud data of dynamic targets; Perform pre-processing, surface reconstruction, mesh optimization and post-processing on point cloud data; The same space is scanned repeatedly to complete the scene construction of the three-dimensional static space.
4. The closed-loop virtual simulation test method for unmanned equipment according to claim 3, characterized in that: Perform pre-processing, surface reconstruction, mesh optimization, and post-processing on point cloud data. The specific steps include: Statistical filtering is used to remove noise points, multi-view point cloud registration is performed based on feature matching or iterative closest point algorithm, and voxel filtering is used to streamline point cloud data to improve processing efficiency. Use Delaunay triangulation to construct the initial triangular mesh, fit the smooth surface and extract the triangular mesh through the moving least squares method, or use Poisson reconstruction to solve the Poisson equation to generate an implicit surface and then convert it into a triangular mesh; Use the Laplace smoothing algorithm to adjust vertex positions to smooth the surface, and use the edge contraction algorithm to simplify the mesh and repair holes and topological errors; Map photogrammetric or hand-painted textures onto mesh models, add lighting and material effects using graphics rendering libraries, and generate realistic 3D mesh models.
5. The closed-loop virtual simulation test method for unmanned equipment according to claim 4, characterized in that: Input the static obstacle coordinate information and dynamic obstacle trajectory information into the 3D static scene and match it with the existing map data. Obtain the precise location and attribute information of roads, buildings, and traffic signs based on high-precision map data, lidar point cloud data, and camera image data collected by sensors; By matching the vehicle's position with a high-precision map, the coordinate information of static obstacles around the vehicle is determined; The static obstacle coordinate information is input into the 3D static scene and matched with the existing 3D static map.
6. The closed-loop virtual simulation test method for unmanned equipment according to claim 5, characterized in that: Input the static obstacle coordinate information and dynamic obstacle trajectory information into the 3D static scene and match it with the existing map data. Process sensor data in real time and build a multi-hypothesis tracking model for dynamic obstacle tracking, including: At time step t, the sensor observation set Z t ={z1, z2, ..., z Nt } and the target hypothesis set H t-1 Make an association; Calculate the Mahalanobis distance d(z,h)=[(zz^) T S -1 (zz^)] 1 / 2 Evaluate the match between observations and hypotheses and construct a data association matrix, where z^ is the predicted observation of the target hypothesis h and S is the covariance matrix; Generate target-observation association hypothesis combinations based on branch-and-bound algorithm; Using the state transfer equation x t =F t x t -1+v t and observation model z t =H t x t +w t Perform status updates and predictions, where F t is the state transfer matrix, v t is the process noise, H t is the observation matrix, w t is the observation noise; According to Bayes' rule, the posterior probability of each hypothesis P(h|Z 1:t )∝P(Z t |h)P(h|Z 1:t-1 ); Retain the high confidence hypothesis set H through pruning strategy t As a result of tracking; Predict the trajectory, velocity and acceleration information of dynamic obstacles; Input dynamic obstacle trajectory information into the 3D static scene.
7. The closed-loop virtual simulation test method for unmanned equipment according to claim 6, characterized in that: After receiving the equipment status information, the simulation software backend module will update the 3D simulation scene synchronously and use NeRF technology to complete or synthesize new perspectives. The simulation software backend module is the basic platform for simulation testing, which is used to integrate the algorithm under test, three-dimensional static space, equipment status information and static environment map.
8. A virtual simulation test closed-loop system for unmanned equipment, characterized in that: The closed-loop virtual simulation test method for unmanned equipment according to any one of claims 1 to 7 is used for operation.