A method and system for autonomous driving simulation testing based on fault perception
By acquiring ground truth data and sensor perception data from autonomous vehicles, and using Gaussian mixture models and Markov chains to generate realistic perception fault data, the problem of being unable to identify potential dangerous behaviors in autonomous driving simulation testing is solved, thereby improving the safety and comfort of autonomous driving.
Patent Information
- Application Number
- CN202311263416.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-27
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-09-27
AI Technical Summary
Current autonomous driving simulation tests cannot effectively identify potentially dangerous behaviors, resulting in poor safety and comfort.
By acquiring ground truth data and sensor perception data from autonomous vehicles, we can identify perception fault scenarios, extract perception fault data and their corresponding ground truth data, and generate realistic perception fault data using Gaussian mixture models and Markov chains for autonomous driving simulation testing.
It can identify potentially dangerous behaviors of autonomous driving under conditions of perception limitations, thus improving the safety and comfort of autonomous driving.
Smart Images

Figure CN117236050B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of autonomous driving, and more specifically, to an autonomous driving simulation testing method and system based on perceived faults. Background Technology
[0002] Autonomous vehicles perceive their surroundings and conditions using various onboard sensors such as cameras, LiDAR, millimeter-wave radar, and ultrasonic radar. Based on the information obtained about roads, traffic, vehicle position, and obstacles, they autonomously analyze and make judgments, plan paths, and control vehicle behavior. To ensure the reliability and safety of autonomous driving functions, autonomous driving algorithms need to undergo thorough testing before mass production and market launch. Simulation testing has advantages over real-vehicle testing in terms of lower cost, faster testing speed, and higher scenario efficiency, and is therefore widely used for software function verification. Currently, autonomous driving simulation testing mostly uses ideal sensor models, primarily for verifying decision-making, planning, and control algorithms. Analysis of problems encountered in real-vehicle testing reveals that perception issues account for 40% of the problems. Furthermore, the information perceived by sensors is rarely absolutely stable and reliable. Therefore, establishing perception failure scenarios for testing is crucial for promoting rapid algorithm iteration and optimizing the mass production of autonomous vehicles.
[0003] Existing technology provides an evaluation method for autonomous driving test scenarios based on perception defects, including: acquiring autonomous driving scenarios for testing; calculating the degree of road risk exposure, accident risk level, and accident severity based on static and dynamic traffic elements of the autonomous driving scenario; classifying the autonomous driving scenario according to the perception level of different types of onboard sensors, including full perception, partial perception, or no perception, and rating the risk level of each type of autonomous driving scenario to obtain the perception defect risk level; and establishing an evaluation model based on the obtained risk exposure level, accident risk level, accident severity, and perception defect risk level to identify key specific scenarios for autonomous driving scenario risk testing evaluation. This existing technology uses extracted key scenarios for risk testing evaluation but does not introduce perception defect scenarios, thus failing to effectively identify potential dangerous behaviors and consequently failing to ensure the safety and comfort of autonomous driving. Summary of the Invention
[0004] One objective of this invention is to provide an autonomous driving simulation test method based on perceived faults, so as to overcome the shortcomings of the prior art in autonomous driving simulation tests, which cannot effectively identify potential dangerous behaviors, resulting in low safety and poor comfort of autonomous driving; the second objective is to provide an autonomous driving simulation test system based on perceived faults.
[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides an autonomous driving simulation testing method based on perceived faults, comprising: Acquire map data, ground truth data, and sensor perception data during the operation of autonomous vehicles; Based on the true value data and sensor perception data, the perception fault scenario is determined, and the perception fault data and its corresponding true value data are extracted. The perceived fault data is processed to obtain the type of perceived fault data and the subtypes of each type of perceived fault data; Based on the ground truth data and map data corresponding to the perceived fault data, scene labels are set; The subtypes of the perceived fault data are matched with scene labels to obtain the fault-scene mapping relationship; Based on the fault-scenario mapping relationship, real perceived fault data is generated for the corresponding real-value scenarios in the preset autonomous driving scenario library. Autonomous driving simulation tests are conducted based on the true scenarios and corresponding real-world perceived fault data in the autonomous driving scenario library.
[0006] Preferably, the truth data of the autonomous vehicle during operation is obtained by using the truth system installed on the autonomous vehicle; and the sensor perception data of the autonomous vehicle during operation is obtained by using the on-board sensors installed on the autonomous vehicle.
[0007] Preferably, the vehicle-mounted sensor includes one or more of the following: camera, lidar, ultrasonic radar, and millimeter-wave radar.
[0008] Preferably, the specific method for determining the perceived fault scenario based on truth data and sensor perception data, and extracting the perceived fault data and its corresponding truth data is as follows: Synchronize the true value data with the sensor-sensed data in time frames; The true data and sensor-sensed data of the same time frame are compared to obtain the comparison deviation; If the comparison deviation is greater than the preset deviation threshold for several consecutive time frames, then the sensor sensing data between the start time frame and the end time frame of the comparison deviation that is greater than the preset deviation threshold will be used as the sensing fault scenario. Extract the start time frame of the perceived fault scene Seconds until the end time frame The sensor data corresponding to a time interval of seconds is used as fault detection data;
[0009] Extract the ground truth data that is in the same time period as the perceived fault data, and use it as the ground truth data corresponding to the perceived fault data.
[0010] Preferably, the perceived fault data is processed to obtain the types of perceived fault data, including type identification fault, motion state fault data, and existence identification fault.
[0011] Preferably, a Gaussian mixture model is used to cluster the various types of perceived fault data to obtain subtypes of each type of perceived fault data, specifically including: A Gaussian mixture model was used to cluster type recognition faults to obtain subtypes of type recognition faults, including pedestrian misidentification, two-wheeled vehicle misidentification, small vehicle misidentification, large vehicle misidentification, lane line misidentification, and drivable area boundary point type error. A Gaussian mixture model was used to cluster the motion state fault data to obtain subtypes of motion state fault data, including target lateral and longitudinal acceleration errors, target lateral and longitudinal velocity errors, target lateral and longitudinal distance errors, target heading angle errors, target geometric dimension recognition errors, and motion state errors of drivable area boundary points. A Gaussian mixture model was used to cluster existence identification faults to obtain subtypes of existence identification faults, including target omission, false targets, and confidence bias.
[0012] Gaussian mixture models were used to cluster the various types of perceived fault data, and the Gaussian mixture models are shown in the following equations:
[0013] in ,in This represents the number of Gaussian kernels in the Gaussian mixture model, i.e., the number of subtypes divided by clustering under any type of perceived fault data; Indicates the first The probability of each Gaussian component occurring in the model. , They represent the first The mean and variance matrices corresponding to each Gaussian component. Input quantity A parameter set for any type of perceived fault data;
[0014] The fitting is performed using the Expectation Maximization (EM) algorithm with bounded parameter range, where the E-step calculates the Gaussian components. For observation Data responsiveness:
[0015]
[0016] in, Let be the conditional probability, representing the data point. Belongs to the The probability of a Gaussian component;
[0017] M-steps calculate the model parameters for the new iteration:
[0018]
[0019]
[0020] in, and The expression is:
[0021]
[0022]
[0023] In the above formula, and They respectively represent the bounded range The first and second moments of Gaussian within the timeframe; and The possible values are: , ;
[0024] The convergence condition during training is:
[0025] To determine the number of Gaussian kernels, the Bayesian Information Criterion (BIC) is used as the criterion. Optimize:
[0026]
[0027] in, Indicates the number of data groups in the dataset;
[0028] Choose the minimum K value when the BIC value is low as the number of Gaussian kernels in the Gaussian mixture model:
[0029] The same method is used to calculate subtypes of other types of perceived fault data.
[0030] Preferably, the scene tags include dynamic traffic scene tags and static feature scene tags; The dynamic traffic scene tags and static feature scene tags are classified to obtain dynamic traffic scene sub-tags and static feature scene sub-tags respectively; The dynamic traffic scene sub-tags include target object type scene sub-tags, target object behavior scene sub-tags, and vehicle behavior scene sub-tags; The sub-labels for the vehicle's behavior scenarios include following another vehicle, changing lanes, entering the ramp from the main road, entering the main road from the ramp, and passing through intersections; The static feature class scene sub-labels include obstacle type scene sub-labels, curve curvature scene sub-labels, road elevation scene sub-labels, road superelevation scene sub-labels, road type scene sub-labels, and road facility scene sub-labels.
[0031] Preferably, based on the fault-scenario mapping relationship, a Markov chain is used to generate real perceived fault data corresponding to the ground truth scenarios in the preset autonomous driving scenario library.
[0032] The process of generating real-world perceived fault data using Markov chains to correspond to ground truth scenarios in a pre-defined autonomous driving scenario library involves two steps: 1) the transition probability between the scenario state at the current time and the previous time; and 2) the emission probability between the real scenario state and the features of the perceived fault data. The first step is to construct the transition probabilities. Markov chains describe state transitions and are generally used to describe the dynamic transition characteristics of a system, while general stochastic processes describe the relationship between states and observation sequences. The operation of an autonomous vehicle is viewed as consisting of perceived faults and the current scene, with perceived faults set as hidden states. The current scene is set to observation mode. The test scenario is described using a Markov chain; the initial state of the Markov chain is set as the initial state of the scenario. State transition probability matrix ,but:
[0033]
[0034] in, Representation of features Transfer to features The number of times;
[0035] The second step is to construct the transmission probability, and the set of hidden states for detecting faults is as follows: , Indicates that the perceived fault is in the first The hidden state at time step; the set of observed states of the scene sequence is , Indicates each scene in the first... The observed state at each time step; the constructed emission probability is represented by a Gaussian mixture model;
[0036] Based on steps one and two, the timing and type of fault injection are generated:
[0037] )
[0038] Solving the above equation yields the timing of the sensing fault injection. and the optimal sequence for detecting faults .
[0039] Preferably, the preset autonomous driving scenario library includes a traffic accident scenario library, a standard regulation scenario library, a natural driving scenario library, and an enterprise functional testing requirement scenario library.
[0040] Preferably, the specific method for conducting autonomous driving simulation testing based on the true scenarios and corresponding real-world perceived fault data in the autonomous driving scenario library is as follows: The true scenarios and corresponding real perceived fault data from the autonomous driving scenario library are input into the autonomous driving system for simulation testing to obtain test results. The test results are judged based on the true scenarios to evaluate whether the autonomous driving perception fusion algorithm can filter perceived fault data. For cases where the confidence level is lower than the preset confidence threshold, the algorithm for planning control, predictive decision-making, human-machine interface, etc., can adopt conservative scenario strategies to ensure the safety and comfort of autonomous driving.
[0041] The present invention also provides an autonomous driving simulation test system based on perceived faults, for implementing the above-mentioned method, comprising: The data acquisition module is used to acquire map data, ground truth data, and sensor perception data during the operation of autonomous vehicles; The fault perception data extraction module is used to determine the fault perception scenario based on the true value data and sensor perception data, and to extract the fault perception data and its corresponding true value data. The fault data processing module is used to process the fault data to obtain the type of fault data and the subtypes of each type of fault data. The scene label setting module is used to set scene labels based on the ground truth data and map data corresponding to the perceived fault data; The matching module is used to match the subtypes of the perceived fault data with scene labels to obtain the fault-scene mapping relationship; The real-perception fault data generation module is used to generate real-perception fault data for the corresponding true-value scenarios in the preset autonomous driving scenario library according to the fault-scenario mapping relationship. The perception fault simulation test module is used to conduct autonomous driving simulation tests based on the true scenarios and corresponding real perception fault data in the autonomous driving scenario library.
[0042] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention identifies perceived fault scenarios by acquiring ground truth data and sensor perception data from autonomous vehicles during operation, extracting the perceived fault data and its corresponding ground truth data; processing the perceived fault data to obtain the types of perceived fault data and subtypes of each type; setting scene labels using the corresponding ground truth data and map data; then matching the subtypes of perceived fault data with the scene labels to obtain a fault-scene mapping relationship, and generating realistic perceived fault data corresponding to the ground truth scenarios in a pre-defined autonomous driving scenario library based on the fault-scene mapping relationship; finally, using the ground truth scenarios in the autonomous driving scenario library and the corresponding realistic perceived fault data, conducting autonomous driving simulation tests. This invention generates realistic perceived fault data, which, together with ground truth scenarios, enables the identification of potential dangerous behaviors of autonomous driving under perception limitations, improving the safety and comfort of autonomous driving. Attached Figure Description
[0043] Figure 1 This is a flowchart of an autonomous driving simulation test method based on perceived faults, as described in Example 1.
[0044] Figure 2 This is a flowchart of an autonomous driving simulation test method based on perceived faults, as described in Example 2.
[0045] Figure 3 This is a schematic diagram of the subtypes of type identification faults described in Example 2.
[0046] Figure 4 This is a schematic diagram of the subtypes of motion state fault data described in Example 2.
[0047] Figure 5 This is a schematic diagram of the subtype of existence identification fault described in Example 2.
[0048] Figure 6 This is a schematic diagram of the dynamic traffic scene sub-label described in Example 2.
[0049] Figure 7 This is a schematic diagram of the static feature class scene sub-labels described in Example 2.
[0050] Figure 8 This is a schematic diagram of the structure of an autonomous driving simulation test system based on perceived faults, as described in Example 3. Detailed Implementation
[0051] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.
[0052] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0053] Example 1 This embodiment provides an autonomous driving simulation testing method based on perceived faults, such as... Figure 1 As shown, it includes: S1: Acquire map data, ground truth data, and sensor perception data during the operation of autonomous vehicles; S2: Based on the true value data and sensor perception data, determine the perception fault scenario, and extract the perception fault data and its corresponding true value data; S3: Process the perceived fault data to obtain the type of perceived fault data and the subtypes of each type of perceived fault data; S4: Set scene labels based on the true value data and map data corresponding to the perceived fault data; S5: Match the subtypes of the perceived fault data with the scene labels to obtain the fault-scene mapping relationship; S6: Based on the fault-scenario mapping relationship, generate real perceived fault data for the corresponding real-value scenarios in the preset autonomous driving scenario library; S7: Conduct autonomous driving simulation tests based on the true scenarios and corresponding real-world perceived fault data in the autonomous driving scenario library.
[0054] In its specific implementation, this embodiment determines the perceived fault scenario by acquiring ground truth data and sensor perception data during the operation of the autonomous vehicle, and extracts the perceived fault data and its corresponding ground truth data. The perceived fault data is processed to obtain the type of perceived fault data and the subtypes of each type. Scene labels are set using the corresponding ground truth data and map data. Then, the subtypes of the perceived fault data are matched with the scene labels to obtain a fault-scene mapping relationship. Based on this mapping relationship, real perceived fault data is generated for the ground truth scenarios in the preset autonomous driving scenario library. Finally, autonomous driving simulation tests are conducted using the ground truth scenarios and corresponding real perceived fault data from the autonomous driving scenario library. This invention generates real perceived fault data, which, together with the ground truth scenarios, is used in autonomous driving simulation tests. This enables the identification of potential dangerous behaviors of autonomous driving under perception limitations, improving the safety and comfort of autonomous driving.
[0055] Example 2 This embodiment provides an autonomous driving simulation testing method based on perceived faults, such as... Figure 2As shown, it includes: Acquire map data, ground truth data, and sensor perception data from autonomous vehicles during operation; specifically: Obtain truth data of autonomous vehicles during operation using the truth system onboard the autonomous vehicle; Utilize onboard sensors installed on autonomous vehicles to acquire sensor perception data during the operation of autonomous vehicles; The vehicle-mounted sensors include one or more of the following: cameras, lidar, ultrasonic radar, and millimeter-wave radar.
[0056] Based on the true data and sensor perception data, the perception fault scenario is determined, and the perception fault data and its corresponding true data are extracted; the specific method is as follows: Synchronize the true value data with the sensor-sensed data in time frames; The true data and sensor-sensed data of the same time frame are compared to obtain the comparison deviation; If the comparison deviation is greater than the preset deviation threshold for several consecutive time frames, then the sensor sensing data between the start time frame and the end time frame of the comparison deviation that is greater than the preset deviation threshold will be used as the sensing fault scenario. Extract the start time frame of the perceived fault scene Seconds until the end time frame The sensor data corresponding to a time interval of seconds is used as fault detection data;
[0057] The sensor data for the remaining time period is considered normal sensing data.
[0058] In this embodiment, if the comparison deviation is greater than 10% for 20 consecutive time frames, the sensor perception data between the start time frame and the end time frame is considered as a perception fault scenario. The sensor perception data corresponding to the time period from 7 seconds before the start time frame to 8 seconds after the end time frame of the perceived fault scenario is extracted as the perceived fault data.
[0059] The perceived fault data is processed to obtain the types of perceived fault data, including type identification fault, motion state fault data, and existence identification fault. Gaussian mixture models are used to cluster the various types of perceived fault data to obtain subtypes for each type of perceived fault data; the Gaussian mixture models are shown in the following equations:
[0060] in ,in This represents the number of Gaussian kernels in the Gaussian mixture model, i.e., the number of subtypes divided by clustering under any type of perceived fault data; Indicates the first The probability of each Gaussian component occurring in the model. , They represent the first The mean and variance matrices corresponding to each Gaussian component. Input quantity A parameter set for any type of perceived fault data;
[0061] The fitting is performed using the Expectation Maximization (EM) algorithm with bounded parameter range, where the E-step calculates the Gaussian components. For observation Data responsiveness:
[0062]
[0063] in, Let be the conditional probability, representing the data point. Belongs to the The probability of a Gaussian component;
[0064] M-steps calculate the model parameters for the new iteration:
[0065]
[0066]
[0067] in, and The expression is:
[0068]
[0069]
[0070] In the above formula, and They respectively represent the bounded range The first and second moments of Gaussian within the timeframe; and The possible values are: , ;
[0071] The convergence condition during training is:
[0072] To determine the number of Gaussian kernels, the Bayesian Information Criterion (BIC) is used as the criterion. Optimize:
[0073]
[0074] in, Indicates the number of data groups in the dataset;
[0075] Choose the minimum K value when the BIC value is low as the number of Gaussian kernels in the Gaussian mixture model:
[0076] The same method was used to calculate subtypes of other types of perceived fault data; like Figure 3 As shown, a Gaussian mixture model is used to cluster type recognition faults to obtain subtypes of type recognition faults, including pedestrian misidentification, two-wheeled vehicle misidentification, small vehicle misidentification, large vehicle misidentification, lane line misidentification, and drivable area border point type error. like Figure 4 As shown, a Gaussian mixture model is used to cluster the motion state fault data to obtain subtypes of motion state fault data, including target lateral and longitudinal acceleration errors, target lateral and longitudinal velocity errors, target lateral and longitudinal distance errors, target heading angle errors, target geometric size identification errors, and motion state errors of drivable area boundary points. like Figure 5 As shown, a Gaussian mixture model is used to cluster existence identification faults to obtain subtypes of existence identification faults, including target omission, false targets, and confidence bias.
[0077] Based on the true value data and map data corresponding to the perceived fault data, scene labels are set; The scene tags include dynamic traffic scene tags and static feature scene tags; The dynamic traffic scene tags and static feature scene tags are classified to obtain dynamic traffic scene sub-tags and static feature scene sub-tags respectively; like Figure 6 As shown, the dynamic traffic scene sub-tags include target object type scene sub-tags, target object behavior scene sub-tags, and vehicle behavior scene sub-tags; The sub-labels for the vehicle's behavior scenarios include following another vehicle, changing lanes, entering the ramp from the main road, entering the main road from the ramp, and passing through intersections; like Figure 7As shown, the static feature class scene sub-labels include obstacle type scene sub-labels, curve curvature scene sub-labels, road elevation scene sub-labels, road superelevation scene sub-labels, road type scene sub-labels, and road facility scene sub-labels; The subtypes of the perceived fault data are matched with scene labels to obtain the fault-scene mapping relationship; Based on the fault-scenario mapping relationship, Markov chains are used to generate real perceived fault data corresponding to the ground truth scenarios in the preset autonomous driving scenario library. It includes two steps: 1) the transition probability between the scene state at the current moment and the previous moment; 2) the emission probability between the real scene state and the perceived fault data features; The first step is to construct the transition probabilities. Markov chains describe state transitions and are generally used to describe the dynamic transition characteristics of a system, while general stochastic processes describe the relationship between states and observation sequences. The operation of an autonomous vehicle is viewed as consisting of perceived faults and the current scene, with perceived faults set as hidden states. The current scene is set to observation mode. The test scenario is described using a Markov chain; the initial state of the Markov chain is set as the initial state of the scenario. State transition probability matrix ,but:
[0078]
[0079] in, Representation of features Transfer to features The number of times;
[0080] The second step is to construct the transmission probability, and the set of hidden states for detecting faults is as follows: , Indicates that the perceived fault is in the first The hidden state at time step; the set of observed states of the scene sequence is , Indicates each scene in the first... The observed state at each time step; the constructed emission probability is represented by a Gaussian mixture model;
[0081] Based on steps one and two, the timing and type of fault injection are generated:
[0082] )
[0083] Solving the above equation yields the timing of the sensing fault injection. and the optimal sequence for detecting faults
[0084] Based on the true scenarios and corresponding real-world perceived fault data in the aforementioned autonomous driving scenario library, autonomous driving simulation tests are conducted; specifically: The preset autonomous driving scenario library includes a traffic accident scenario library, a standard and regulation scenario library, a natural driving scenario library, and an enterprise functional testing requirement scenario library.
[0085] The true scenarios and corresponding real perceived fault data from the autonomous driving scenario library are input into the autonomous driving system for simulation testing to obtain test results. The test results are judged based on the true scenarios to evaluate whether the autonomous driving perception fusion algorithm can filter perceived fault data. For cases where the confidence level is lower than the preset confidence threshold, the algorithm for planning control, predictive decision-making, human-machine interface, etc., can adopt conservative scenario strategies to ensure the safety and comfort of autonomous driving.
[0086] Example 3 This embodiment provides an autonomous driving simulation test system based on perceived faults, used to implement the method described in Embodiment 1 or 2, such as... Figure 8 As shown, it includes: The data acquisition module is used to acquire map data, ground truth data, and sensor perception data during the operation of autonomous vehicles; The fault perception data extraction module is used to determine the fault perception scenario based on the true value data and sensor perception data, and to extract the fault perception data and its corresponding true value data. The fault data processing module is used to process the fault data to obtain the type of fault data and the subtypes of each type of fault data. The scene label setting module is used to set scene labels based on the ground truth data and map data corresponding to the perceived fault data; The matching module is used to match the subtypes of the perceived fault data with scene labels to obtain the fault-scene mapping relationship; The real-perception fault data generation module is used to generate real-perception fault data for the corresponding true-value scenarios in the preset autonomous driving scenario library according to the fault-scenario mapping relationship. The perception fault simulation test module is used to conduct autonomous driving simulation tests based on the true scenarios and corresponding real perception fault data in the autonomous driving scenario library.
[0087] The same or similar labels correspond to the same or similar parts; The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. An autonomous driving simulation testing method based on perceived faults, characterized in that, include: Acquire map data, ground truth data, and sensor perception data during the operation of autonomous vehicles; Based on the true value data and sensor perception data, the perception fault scenario is determined, and the perception fault data and its corresponding true value data are extracted. The process of determining the perceived fault scenario based on the true data and sensor-sensed data includes: Synchronize the true value data with the sensor-sensed data in time frames; The true data and sensor-sensed data of the same time frame are compared to obtain the comparison deviation; If the comparison deviation is greater than the preset deviation threshold for several consecutive time frames, then the sensor sensing data between the start time frame and the end time frame of the comparison deviation that is greater than the preset deviation threshold will be used as the sensing fault scenario. The perceived fault data is processed to obtain the type of perceived fault data and the subtypes of each type of perceived fault data; The types of perceived fault data include type identification faults, motion state fault data, and existence identification faults. A Gaussian mixture model is used to cluster type recognition faults to obtain subtypes of type recognition faults; A Gaussian mixture model is used to cluster motion state fault data to obtain subtypes of motion state fault data. A Gaussian mixture model is used to cluster existence identification faults to obtain subtypes of existence identification faults; Based on the ground truth data and map data corresponding to the perceived fault data, scene labels are set; The subtypes of the perceived fault data are matched with scene labels to obtain the fault-scene mapping relationship; Based on the fault-scenario mapping relationship, Markov chains are used to generate real perceived fault data corresponding to the ground truth scenarios in the preset autonomous driving scenario library. Autonomous driving simulation tests are conducted based on the true scenarios and corresponding real-world perceived fault data in the autonomous driving scenario library.
2. The autonomous driving simulation test method based on perceived faults according to claim 1, characterized in that, The system utilizes the truth system mounted on the autonomous vehicle to obtain truth data during the operation of the autonomous vehicle; it also utilizes the onboard sensors mounted on the autonomous vehicle to obtain sensor perception data during the operation of the autonomous vehicle.
3. The autonomous driving simulation test method based on perceived faults according to claim 2, characterized in that, The vehicle-mounted sensors include one or more of the following: cameras, lidar, ultrasonic radar, and millimeter-wave radar.
4. The autonomous driving simulation test method based on perceived faults according to claim 1, characterized in that, The specific method for extracting perceived fault data and its corresponding truth data is as follows: Extract the start time frame of the perceived fault scene Seconds until the end time frame The sensor data corresponding to a time interval of seconds is used as fault detection data; Extract the ground truth data that is in the same time period as the perceived fault data, and use it as the ground truth data corresponding to the perceived fault data.
5. The autonomous driving simulation test method based on perceived faults according to claim 1, characterized in that, The subtypes of the type of fault identification include pedestrian misidentification, two-wheeled vehicle misidentification, car misidentification, large vehicle misidentification, lane line misidentification, and drivable area border point type error; The subtypes of motion state fault data include target lateral and longitudinal acceleration errors, target lateral and longitudinal velocity errors, target lateral and longitudinal distance errors, target heading angle errors, target geometric dimension identification errors, and motion state errors of drivable area boundary points; The subtypes of existence identification failures include target omission, false targets, and confidence bias.
6. The autonomous driving simulation test method based on perceived faults according to claim 1, characterized in that, The scene tags include dynamic traffic scene tags and static feature scene tags; The dynamic traffic scene tags and static feature scene tags are classified to obtain dynamic traffic scene sub-tags and static feature scene sub-tags respectively; The dynamic traffic scene sub-tags include target object type scene sub-tags, target object behavior scene sub-tags, and vehicle behavior scene sub-tags; The static feature class scene sub-labels include obstacle type scene sub-labels, curve curvature scene sub-labels, road elevation scene sub-labels, road superelevation scene sub-labels, road type scene sub-labels, and road facility scene sub-labels.
7. The autonomous driving simulation test method based on perceived faults according to claim 1, characterized in that, The preset autonomous driving scenario library includes a traffic accident scenario library, a standard and regulation scenario library, a natural driving scenario library, and an enterprise functional testing requirement scenario library.
8. An autonomous driving simulation testing system based on perceived faults, used to implement the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire map data, ground truth data, and sensor perception data during the operation of autonomous vehicles; The fault perception data extraction module is used to determine the fault perception scenario based on the true value data and sensor perception data, and to extract the fault perception data and its corresponding true value data. The fault data processing module is used to process the fault data to obtain the type of fault data and the subtypes of each type of fault data. The scene label setting module is used to set scene labels based on the ground truth data and map data corresponding to the perceived fault data; The matching module is used to match the subtypes of the perceived fault data with scene labels to obtain the fault-scene mapping relationship; The real-perception fault data generation module is used to generate real-perception fault data for the corresponding true-value scenarios in the preset autonomous driving scenario library according to the fault-scenario mapping relationship. The perception fault simulation test module is used to conduct autonomous driving simulation tests based on the true scenarios and corresponding real perception fault data in the autonomous driving scenario library.
Citation Information
Patent Citations
Optimized in-loop simulation test method for automobile auxiliary driving controller
CN112578683A
Simulation test system and method for automatic driving
CN115016323A
