Automobile automatic driving test method for simulating complex environment

Abstract

The invention provides an automobile automatic driving test method for simulating a complex environment, and the method comprises the following steps: constructing a simulation test scene and a test simulation body for testing an automobile, and mapping the test simulation body into the simulation test scene; querying accident-prone nodes in the simulation test scene, marking one of the accident-prone nodes as a false node, and marking the other nodes as accident nodes; common accidents are injected into the accident nodes for simulation, and false accident information of false nodes is generated at the same time; performing an automatic driving test on the test simulation body in the simulation test scene, and sending false accident information to the test simulation body; acquiring decision data passing through the false node and the accident node in the test process of the test simulation body, and acquiring a test result based on the decision data; by testing whether the tested vehicle blindly follows and identifying common accidents, the automatic driving performance of the vehicle can be fully tested, so that a decision strategy can be optimized in time.

Description

Technical Field

[0001] This invention relates to the field of automotive driving technology, and in particular to a method for testing autonomous driving of automobiles in a simulated complex environment. Background Technology

[0002] With the continuous increase in the number of cars, the urban road network is becoming increasingly complex, traffic flow density is surging, and road complexity is increasing exponentially. Autonomous driving systems need to operate safely in complex environments such as diverse road conditions and dynamic traffic flow. Therefore, the testing phase needs to cover extreme working conditions and sudden scenarios. Currently, most commonly used testing methods rely on real vehicle road testing or static simulation. Real vehicle road testing requires a large investment of manpower, material resources, and site resources, resulting in high costs and long testing cycles. Furthermore, the scenarios during testing cannot reproduce complex and dangerous environments such as chain-reaction rear-end collisions or accidents caused by severe weather. Static simulation, on the other hand, constructs environments based on preset parameters, which easily overlooks the dynamics of the environment and the interaction between different elements, leading to insufficient assessment of the adaptability to complex scenarios.

[0003] V2X (Vehicle-to-Everything) technology is the core support for autonomous driving. Through multi-dimensional communication links such as vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I), it enables real-time interaction and sharing of traffic information, breaks through the perception blind spots of a single vehicle sensor, and significantly improves the vehicle's perception range and decision-making timeliness. With the help of V2X networks, vehicles can obtain information such as accidents and traffic congestion ahead in advance, adjust driving strategies in advance, reduce collision risks, and greatly improve decision-making timeliness and driving safety. At present, V2X technology has been gradually applied to autonomous driving test scenarios, but existing tests mostly focus on information transmission efficiency and system response under normal information interaction, and lack in-depth verification of the propagation characteristics of abnormal information in the network and the system response mechanism.

[0004] However, the openness and heterogeneity of V2X networks highlight the risk of spreading false traffic information. For example, malicious nodes or network attacks may inject false alarms such as accidents ahead or lane closures. This kind of information can easily spread rapidly among vehicles, forming an information epidemic, thanks to the broadcast characteristics of V2X. Some autonomous driving systems, lacking effective information verification mechanisms, may blindly follow false information and take actions such as emergency braking and disorderly lane changes without local sensor cross-verification. This not only affects driving efficiency but may also induce secondary traffic accidents. Existing autonomous driving test schemes do not have targeted evaluation processes designed for false information scenarios. They cannot simulate the spread of false information in vehicle groups, nor can they quantify the system's blind following behavior and error correction capabilities. As a result, the safety hazards of autonomous driving systems in V2X information interaction scenarios have not been fully exposed. Summary of the Invention

[0005] In view of this, the present invention proposes a test method for autonomous driving of automobiles that simulates complex environments, which can simultaneously test the vehicle's ability to identify real accidents and fake accidents, thereby improving the autonomous driving performance of automobiles in complex environments.

[0006] The technical solution of this invention is implemented as follows: A method for testing autonomous driving in a simulated complex environment includes the following steps: Step S1: Construct a simulation test scenario and a test simulation object of the test vehicle, and map the test simulation object into the simulation test scenario; Step S2: Query the accident-prone nodes in the simulation test scenario, mark one of the accident-prone nodes as a false node, and mark the others as accident nodes; Step S3: Inject common accidents into the accident nodes to simulate them, and at the same time generate false accident information for false nodes; Step S4: Conduct autonomous driving tests on the test simulator in the simulation test scenario, and send false accident information to the test simulator; Step S5: Collect decision data of the test simulation object during the test of false nodes and accident nodes, and obtain test results based on the decision data.

[0007] Preferably, step S1 includes the following steps: Step S11: Obtain road network data for vehicle testing, construct the basic geographic framework of the 3D simulation test scenario based on the road network data, and set dynamic elements of the scenario; Step S12: Collect the physical parameters and sensor configuration of the test vehicle to construct a test simulation, and load the autonomous driving decision-making algorithm into the test simulation; Step S13: Map the test simulation object to the initial position of the simulation test scene.

[0008] Preferably, the road network data includes lane lines, traffic signs, and intersections; the physical parameters include wheelbase, turning radius, and braking performance; and the sensor configuration includes lidar field of view and camera resolution.

[0009] Preferably, step S2 includes the following specific steps: Step S21: Model the road topology of the simulation test scenario and identify key intersections in the traffic flow; Step S22: Extract historical traffic data of key intersections, construct a feature dataset, and preprocess the feature dataset; Step S23: Use the trained XGBoost model to process the preprocessed feature dataset and obtain the probability of an accident at each key intersection. Step S24: Output the key intersections with an accident probability greater than a preset threshold as accident nodes, and randomly select one of the accident nodes as a false node.

[0010] Preferably, after step S24, a multi-objective optimization algorithm is introduced, with the node risk score difference degree and node spatial distribution density as optimization objectives, and the selection combination of false nodes and accident nodes as optimization variables. The algorithm iterates to generate Pareto optimal solutions, and selects combinations of nodes that are evenly distributed and cover high and medium risk levels.

[0011] Preferably, step S3 includes the following specific steps: Step S31: Based on the different node types, inject common incidents at the incident nodes and set the duration and scope of impact of common incidents; Step S32: Define the event type, location, and severity of false incident information as core attributes; Step S33: Convert the core attributes into information fields through natural language processing and encapsulate them into fake incident information in V2X network information format; Step S34: Use data augmentation techniques to mutate and expand false accident information.

[0012] Preferably, the specific steps of step S34 are as follows: Step S341: Using a pre-trained conditional variational autoencoder, by adjusting its latent space parameters, automatically generate mutation incident information with reasonable biases in key fields; Step S342: Rewrite the text description of the false accident information while preserving its semantics; Step S343: Simulate information sources with different distances, signal strengths, and reliability by randomly perturbing the communication protocol parameters of V2X messages.

[0013] Preferably, step S4 includes the following specific steps: Step S41: Start the simulation test and control the test simulator to pass through false nodes and accident nodes in the simulation test scenario according to the preset path; Step S42: When the test simulator enters the communication range of the fake node, send fake accident information to the test simulator and simulate network latency and packet loss.

[0014] Preferably, the specific steps for step S5, which involves collecting decision data from sham nodes during the testing process of the test simulation, and obtaining test results based on this decision data, are as follows: Step S51: After receiving false accident information, the test simulation determines whether there is a passable path at the false node. Step S52: When a passable path exists, collect the response actions of the test simulation object and the sensor trigger information; Step S53: If the test simulator does not trigger local sensor verification and directly takes a response action, or if the sensor verification result shows no false accident but still executes the response action, it is determined that the test simulator blindly follows.

[0015] Preferably, step S5 involves collecting decision data from accident nodes during the testing of the test simulation, and obtaining test results based on this decision data. The specific steps are as follows: Step S54: Collect decision data when the test simulation object passes through the accident node; Step S55: Calculate the accident identification accuracy, obstacle avoidance path planning rationality, and time taken to pass through the accident area based on the decision data.

[0016] Compared with the prior art, the beneficial effects of the present invention are: ① By using simulation testing methods, after constructing a test simulation body for the test vehicle, simulation tests are carried out in the simulation test scenario. This can simulate a variety of different scenarios, thereby verifying the autonomous driving performance of the test vehicle, so as to adjust and optimize the control and decision-making strategies of the test vehicle. ② In the simulation test scenario, accident-prone nodes will be identified, and one of them will be marked as a false node. False accident information can be sent to the test simulation object at the false node. The response of the test simulation object after receiving the false accident information will determine whether the test simulation object blindly follows or still makes its own judgment, thus evaluating the autonomous decision-making ability of the test vehicle. ③ In addition to false nodes, accident-prone nodes are accident nodes. Common accidents can be injected at accident nodes. When the test simulation moves to an accident node, it can identify common accidents and make corresponding decisions. Based on the identification and decision, the test vehicle can be subjected to regular autonomous driving tests. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for testing autonomous driving in a complex environment according to the present invention; Figure 2 This is a flowchart of step S1 of a method for testing autonomous driving of a vehicle in a simulated complex environment according to the present invention. Figure 3This is a flowchart of step S2 of a vehicle autonomous driving test method simulating a complex environment according to the present invention; Figure 4 This is a flowchart of step S3 of a vehicle autonomous driving test method simulating a complex environment according to the present invention. Figure 5 This is a flowchart of step S34 of a method for testing autonomous driving of a vehicle in a simulated complex environment according to the present invention. Figure 6 This is a flowchart of step S4 of a vehicle autonomous driving test method simulating a complex environment according to the present invention. Figure 7 This is a flowchart of step S5 of a method for testing autonomous driving of a vehicle in a simulated complex environment according to the present invention. Detailed Implementation

[0019] To better understand the technical content of this invention, a specific embodiment is provided below, and the invention will be further described in conjunction with the accompanying drawings.

[0020] See Figures 1 to 7 The present invention provides a method for testing autonomous driving of automobiles in a simulated complex environment, comprising the following steps: Step S1: Construct a simulation test scenario and a test simulation object of the test vehicle, and map the test simulation object into the simulation test scenario; Step S2: Query the accident-prone nodes in the simulation test scenario, mark one of the accident-prone nodes as a false node, and mark the others as accident nodes; Step S3: Inject common accidents into the accident nodes to simulate them, and at the same time generate false accident information for false nodes; Step S4: Conduct autonomous driving tests on the test simulator in the simulation test scenario, and send false accident information to the test simulator; Step S5: Collect decision data of the test simulation object during the test of false nodes and accident nodes, and obtain test results based on the decision data.

[0021] This invention discloses a method for testing autonomous driving in complex environments. It employs simulation testing to conduct autonomous driving tests on test vehicles. A simulated test scenario is constructed based on actual test conditions, capable of reproducing real-world scenarios. This simulation scenario includes road structure, traffic light patterns, weather changes, and interactions between pedestrians and non-motorized vehicles. Simultaneously, a test simulation entity is constructed based on relevant information about the test vehicle. This entity is used to conduct autonomous driving tests within the simulated test scenario. The test results from the simulation entity within the simulated test scenario can be used to evaluate the autonomous driving performance of the test vehicle. Because the test is conducted through simulation, different scenarios can be changed at any time, and rare but high-risk scenarios, such as chain-reaction collisions, can be reproduced. This allows for the evaluation of the test vehicle's autonomous driving performance in various complex environments, assessing obstacle recognition capabilities and path planning rationality.

[0022] In addition to basic path planning and obstacle recognition and avoidance, autonomous driving testing also includes timely decision-making in the event of a traffic accident ahead to avoid traffic congestion and more serious accidents. Therefore, simulation testing incorporates the test vehicle's response to accidents. Within the simulation test scenario, there are specific nodes prone to accidents, such as critical intersections and tunnel entrances / exits. These nodes are designated as accident-prone nodes, with one node designated as a false node and the others as actual accident nodes. Common accidents, such as vehicle collisions and water accumulation on the ground, can be injected into these accident nodes. When the test simulator passes through these accident nodes, the sensor array on the simulator can identify these common accidents and make corresponding decisions based on the identification results. For example, upon detecting water accumulation ahead, the simulator can slowly decelerate and change lanes safely. The obstacle recognition and decision-making processes allow for the evaluation of the test vehicle's autonomous driving performance in the face of common accidents.

[0023] Additionally, one of the accident-prone nodes is designated as a spurious node. The spurious node's role is to send false accident information to the test simulator to test whether the simulator blindly follows the instructions. After identifying a spurious node, false accident information is generated for that node. This false accident information is sent in the form of messages or data packets, containing details of the accident occurring at the spurious node, such as a two-car collision in the leftmost lane. When the test simulator conducts autonomous driving tests in the simulation scenario, it will pass through both accident nodes and spurious nodes. When the test simulator enters the range of a spurious node, false accident information is sent to it. If the test vehicle responds directly upon receiving false accident information, it indicates that it is blindly following the information. The correct response is for the test simulator to interpret the false accident information, drive on a route away from the false accident scene, and use its own sensor array to determine whether a false accident has occurred at the false node. If a decision is made blindly without prior assessment, it indicates that the autonomous driving performance of the test vehicle needs further optimization. Finally, the autonomous driving test results of the test vehicle can be obtained based on the decision data of the test simulator at the false and accident nodes, thereby guiding the autonomous driving strategy of the test vehicle.

[0024] Preferably, step S1 includes the following steps: Step S11: Obtain road network data for vehicle testing, construct the basic geographic framework of the 3D simulation test scenario based on the road network data, and set dynamic elements of the scenario; Step S12: Collect the physical parameters and sensor configuration of the test vehicle to construct a test simulation, and load the autonomous driving decision-making algorithm into the test simulation; Step S13: Map the test simulation object to the initial position of the simulation test scene.

[0025] When constructing a simulation test scenario, it is necessary to obtain actual road network data. First, determine the road network data of roads commonly used for vehicle testing, including lane lines, traffic signs, and intersections. Then, construct a three-dimensional simulation test scenario containing multiple lanes, traffic signs at different locations, and a different number of intersections based on the road network data. At the same time, dynamic elements of the scenario will also be added to the three-dimensional simulation test scenario, such as weather changes, sunny or rainy days. In addition, dynamic elements such as virtual vehicles in other lanes, non-motorized vehicles and pedestrians on sidewalks can also be added.

[0026] When constructing a test simulation, physical parameters of the actual test vehicle can be collected first, including wheelbase, turning radius, and braking performance. Then, based on the physical parameters, the overall framework of the test vehicle is built to form a 1:1 twin of the test vehicle. Sensor configurations, including LiDAR field of view and camera resolution, are collected from the test vehicle and mapped synchronously into the test simulation. This allows the test simulation to accurately simulate the test vehicle's driving, turning, and braking processes during autonomous driving tests.

[0027] Preferably, step S2 includes the following specific steps: Step S21: Model the road topology of the simulation test scenario and identify key intersections in the traffic flow; Step S22: Extract historical traffic data of key intersections, construct a feature dataset, and preprocess the feature dataset; Step S23: Use the trained XGBoost model to process the preprocessed feature dataset and obtain the probability of an accident at each key intersection. Step S24: Output the key intersections with an accident probability greater than a preset threshold as accident nodes, and randomly select one of the accident nodes as a false node.

[0028] In simulation test scenarios, there are several nodes, and different nodes have different probabilities of accidents. Based on past experience, intersections have a higher probability of accidents, such as multi-road intersections and the junctions of tunnels and roads, i.e., tunnel entrances and exits. After modeling the road topology of the simulation test scenario, key intersections can be identified based on traffic flow. However, not all key intersections are prone to accidents; historical traffic data needs to be comprehensively evaluated. After extracting historical traffic data from key intersections, a feature dataset can be constructed. Then, preprocessing such as missing value imputation, outlier removal, and standardization is performed on the feature dataset. The preprocessed feature dataset can then be input into the trained XGBoost model. The OST model is trained and tested using a pre-classified training and test set. After processing the feature dataset, it outputs the probability of an accident at each key intersection. The probability of an accident is compared with a preset threshold. If it is greater than the preset threshold, it indicates that the key intersection is prone to accidents and can be output as an accident node. Among these accident nodes, one needs to be selected as a false node. The selection of the false node does not require too many restrictions, so it can be randomly selected from the accident nodes. In addition, the false node should be avoided before or after all accident nodes, but should be embedded between accident nodes to ensure that the test simulation object receives false accident information for testing only after recognizing common accidents at accident nodes.

[0029] Preferably, after step S24, a multi-objective optimization algorithm is introduced, with the node risk score difference degree and node spatial distribution density as optimization objectives, and the selection combination of false nodes and accident nodes as optimization variables. The algorithm iterates to generate Pareto optimal solutions, and selects combinations of nodes that are evenly distributed and cover high and medium risk levels.

[0030] After identifying accident nodes and false nodes, the NSGA-II multi-objective optimization algorithm is introduced. With the optimization objectives of ensuring safety and improving testing efficiency, and with the node risk score being greater than the score threshold as a constraint, the algorithm iteratively generates Pareto optimal solutions. Then, combinations with evenly distributed nodes covering high and medium risk levels are selected to ensure that the node selection is both representative of safety and testing efficiency.

[0031] Preferably, step S3 includes the following specific steps: Step S31: Based on the different node types, inject common incidents at the incident nodes and set the duration and scope of impact of common incidents; Step S32: Define the event type, location, and severity of false incident information as core attributes; Step S33: Convert the core attributes into information fields through natural language processing and encapsulate them into fake incident information in V2X network information format; Step S34: Use data augmentation techniques to mutate and expand false accident information.

[0032] At the accident node, common accidents need to be injected. Depending on the type of node, the scenario accidents injected will also be different. For example, at the tunnel entrance and exit, a car rear-end collision accident can be injected, and on a potholed road, a water accumulation accident can be injected. Common accidents are injected into the corresponding accident nodes, and the duration and impact range of the accident nodes can be set. For example, the duration of a rear-end collision accident can be set to 30 minutes to 1 hour, and the impact range can be the lane where the accident vehicle is located and the adjacent lanes.

[0033] False accident information is generally transmitted via V2X, so it needs to be encapsulated in the same format. First, the event type, location, and severity of the false accident information are determined, such as a serious multi-vehicle rear-end collision in the rightmost lane at the tunnel entrance. Then, NLP technology is used to convert the core attributes into information fields and encapsulate them into V2X network information format as false accident information. Finally, in order to simulate false alarm scenarios with different credibility and sources, data augmentation technology is used to mutate and expand the false accident information.

[0034] Preferably, the specific steps of step S34 are as follows: Step S341: Using a pre-trained conditional variational autoencoder, by adjusting its latent space parameters, automatically generate mutation incident information with reasonable biases in key fields; Step S342: Rewrite the text description of the false accident information while preserving its semantics; Step S343: Simulate information sources with different distances, signal strengths, and reliability by randomly perturbing the communication protocol parameters of V2X messages.

[0035] Conditional variational autoencoders systematically generate mutated accident information with reasonable deviations in fields such as location coordinates, timestamps, and accident descriptions by adjusting the noise vector and conditional labels in their latent space, such as accident type and severity. At the same time, a text generation model can rewrite the text description of false accident information with semantic preservation to simulate different narrative styles and levels of detail. Finally, through protocol parameter perturbation technology, communication parameters such as the transmission power and frequency of V2X messages are randomly adjusted to simulate information sources with different distances and reliability, thereby realizing the mutation and expansion of false accident information.

[0036] Preferably, step S4 includes the following specific steps: Step S41: Start the simulation test and control the test simulator to pass through false nodes and accident nodes in the simulation test scenario according to the preset path; Step S42: When the test simulator enters the communication range of the fake node, send fake accident information to the test simulator and simulate network latency and packet loss.

[0037] After injecting common accidents and generating false accident information, the testing process of the test simulator can be started. The test simulator can be controlled to drive autonomously in the simulated test scenario. When the test simulator passes an accident node, it will automatically identify and make a response decision. When the test simulator travels to the communication range of the false node, the false accident information can be sent to the test simulator. At the same time, network interference is added during the transmission process, such as increasing network latency and packet loss, to simulate the real communication environment.

[0038] Preferably, the specific steps for step S5, which involves collecting decision data from sham nodes during the testing process of the test simulation, and obtaining test results based on this decision data, are as follows: Step S51: After receiving false accident information, the test simulation determines whether there is a passable path at the false node. Step S52: When a passable path exists, collect the response actions of the test simulation object and the sensor trigger information; Step S53: If the test simulator does not trigger local sensor verification and directly takes a response action, or if the sensor verification result shows no false accident but still executes the response action, it is determined that the test simulator blindly follows.

[0039] When a test simulator enters the communication range of a fake node, it can send false accident information. Upon receiving this information, the test simulator must first parse it to obtain details of the false accident at the fake node, including the specific accident type, the scope of its impact, and whether there is still a passable path. For example, if the fake node is a three-lane road and the false accident is a rear-end collision between two vehicles causing the left two lanes to be inaccessible, the test simulator can still pass through the rightmost lane, which is then the passable path. If the test simulator receives the false accident information, it can directly change its overall travel path to avoid the fake node. When a false node is detected, it indicates that the test simulator may be blindly following. When a passable path exists, the test simulator should travel at a low speed at a distance from the false node and trigger the sensors to identify the false node. If the sensors detect no false incident, the test simulator should switch to normal driving mode. However, if the test simulator still performs its response action, it means that the test simulator is still performing driving control based on the false incident information it received, which can also be judged as blind following. If the test simulator is judged to be blindly following, its identification and autonomous driving strategy need to be optimized to further improve the performance of autonomous driving.

[0040] Preferably, step S5 involves collecting decision data from accident nodes during the testing of the test simulation, and obtaining test results based on this decision data. The specific steps are as follows: Step S54: Collect decision data when the test simulation object passes through the accident node; Step S55: Calculate the accident identification accuracy, obstacle avoidance path planning rationality, and time taken to pass through the accident area based on the decision data.

[0041] When the test simulator passes through an accident node during its driving process, it will make corresponding decisions, such as identifying the type of obstacle, planning a new path based on the scope of the obstacle, and the time taken to pass through the accident area. Since common accident types are relatively common and the countermeasures are relatively complete, by comparing the decision data of the test simulator when passing through the accident node with traditional countermeasures, its ability to deal with common accidents can be evaluated, thereby evaluating the autonomous driving performance of the test vehicle.

[0042] Ultimately, the overall evaluation of the test vehicle can be achieved based on factors such as whether the test vehicle blindly follows others, the accuracy of its identification of common accidents, the rationality of its planned route, and whether the time spent does not increase excessively. If the test vehicle blindly follows others, its decision-making strategy needs to be further optimized so that the test vehicle relies primarily on its own identification, avoiding route changes caused by false information, which could lead to increased user time and more dangerous accidents.

[0043] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for testing autonomous driving in a vehicle under complex environments, characterized in that, Includes the following steps: Step S1: Construct a simulation test scenario and a test simulation object of the test vehicle, and map the test simulation object into the simulation test scenario; Step S2: Query the accident-prone nodes in the simulation test scenario, mark one of the accident-prone nodes as a false node, and mark the others as accident nodes; Step S3: Inject common accidents into the accident nodes to simulate them, and at the same time generate false accident information for false nodes; Step S4: Conduct autonomous driving tests on the test simulator in the simulation test scenario, and send false accident information to the test simulator; Step S5: Collect decision data of the test simulation object during the test of false nodes and accident nodes, and obtain test results based on the decision data.

2. The method for simulating complex environments for testing autonomous driving of vehicles according to claim 1, characterized in that, The specific steps of step S1 include: Step S11: Obtain road network data for vehicle testing, construct the basic geographic framework of the 3D simulation test scenario based on the road network data, and set dynamic elements of the scenario; Step S12: Collect the physical parameters and sensor configuration of the test vehicle to construct a test simulation, and load the autonomous driving decision-making algorithm into the test simulation; Step S13: Map the test simulation object to the initial position of the simulation test scene.

3. The method for simulating complex environments for testing autonomous driving of vehicles according to claim 1, characterized in that, The road network data includes lane lines, traffic signs, and intersections; the physical parameters include wheelbase, turning radius, and braking performance; and the sensor configuration includes lidar field of view and camera resolution.

4. The method for testing autonomous driving in a simulated complex environment according to claim 1, characterized in that, The specific steps of step S2 include: Step S21: Model the road topology of the simulation test scenario and identify key intersections in the traffic flow; Step S22: Extract historical traffic data of key intersections, construct a feature dataset, and preprocess the feature dataset; Step S23: Use the trained XGBoost model to process the preprocessed feature dataset and obtain the probability of an accident at each key intersection. Step S24: Output the key intersections with an accident probability greater than a preset threshold as accident nodes, and randomly select one of the accident nodes as a false node.

5. The method for testing autonomous driving in a simulated complex environment according to claim 4, characterized in that, After step S24, a multi-objective optimization algorithm is introduced, with the node risk score difference and node spatial distribution density as optimization objectives. The selection combination of false nodes and accident nodes is used as optimization variables. The algorithm iterates to generate Pareto optimal solutions, from which combinations with uniform node distribution and coverage of high and medium risk levels are selected.

6. The method for testing autonomous driving in a simulated complex environment according to claim 1, characterized in that, The specific steps of step S3 include: Step S31: Based on the different node types, inject common incidents at the incident nodes and set the duration and scope of impact of common incidents; Step S32: Define the event type, location, and severity of false incident information as core attributes; Step S33: Convert the core attributes into information fields through natural language processing and encapsulate them into fake incident information in V2X network information format; Step S34: Use data augmentation techniques to mutate and expand false accident information.

7. The method for testing autonomous driving of a vehicle in a simulated complex environment according to claim 1, characterized in that, The specific steps of step S34 are as follows: Step S341: Using a pre-trained conditional variational autoencoder, by adjusting its latent space parameters, automatically generate mutation incident information with reasonable biases in key fields; Step S342: Rewrite the text description of the false accident information while preserving its semantics; Step S343: Simulate information sources with different distances, signal strengths, and reliability by randomly perturbing the communication protocol parameters of V2X messages.

8. The method for testing autonomous driving in a simulated complex environment according to claim 1, characterized in that, The specific steps of step S4 include: Step S41: Start the simulation test and control the test simulator to pass through false nodes and accident nodes in the simulation test scenario according to the preset path; Step S42: When the test simulator enters the communication range of the fake node, send fake accident information to the test simulator and simulate network latency and packet loss.

9. The method for testing autonomous driving in a simulated complex environment according to claim 1, characterized in that, Step S5 involves collecting decision data from spoofed nodes during the testing process of the test simulation. The specific steps for obtaining the test results based on this decision data are as follows: Step S51: After receiving false accident information, the test simulation determines whether there is a passable path at the false node. Step S52: When a passable path exists, collect the response actions of the test simulation object and the sensor trigger information; Step S53: If the test simulator does not trigger local sensor verification and directly takes a response action, or if the sensor verification result shows no false accident but still executes the response action, it is determined that the test simulator blindly follows.

10. The method for testing autonomous driving in a simulated complex environment according to claim 1, characterized in that, Step S5 involves collecting decision data from accident nodes during the testing process of the test simulation. The specific steps for obtaining test results based on this decision data are as follows: Step S54: Collect decision data when the test simulation object passes through the accident node; Step S55: Calculate the accident identification accuracy, obstacle avoidance path planning rationality, and time taken to pass through the accident area based on the decision data.

Images