Method and device for verifying an autonomous driving model, controller, vehicle and medium

Abstract

The application provides a kind of automatic driving model verification method, device, controller, vehicle and medium, method includes: in the process of vehicle driving, make the automatic driving model to be verified based on shadow mode operation, collect first environment data and human decision planning result, obtain the risk scene and corresponding data segment of model decision deviation;Human decision planning segment is used as supervision label, and the model is incrementally trained;Control vehicle to run on open road, collect second environment data to obtain model decision planning result and driving scene;Statistical model and driver's feature parameters under each scene category;According to the comparison result of feature parameter, calculate the comprehensive confidence index of model;If the index is greater than the preset threshold, determine that the model passes the verification.Based on this, the application realizes hierarchical, closed-loop model verification method, realizes the quantitative evaluation of the uncertainty of automatic driving model, improves the pertinence and effectiveness of test verification.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, specifically to a verification method, device, controller, vehicle, and medium for an autonomous driving model. Background Technology

[0002] With the development of artificial intelligence technology, the architecture of intelligent driving systems is evolving from the traditional modular architecture of "perception-decision-control" to an end-to-end model architecture. End-to-end autonomous driving models can directly input sensor data and output control signals, simplifying the development process and representing an important direction for the development of intelligent driving.

[0003] Currently, testing of end-to-end autonomous driving models mainly relies on offline dataset evaluation and open-loop / closed-loop simulation testing, which has many shortcomings and restricts industrial application. The core problems are as follows: First, the test fidelity is low. Offline datasets and simulation data cannot replicate real driving scenarios, and even after the model meets offline standards, there may still be safety hazards when it is deployed in the vehicle. Second, there is a lack of targeted testing, which cannot accurately identify the model's defects and potential problems in specific scenarios, increasing the safety risks when deployed in the vehicle. Third, there is no effective solution for evaluating model output. End-to-end models are based on probability, and the output is uncertain. There is currently no optimal evaluation method in the industry. Therefore, existing testing solutions have safety risks and cannot meet actual testing needs, making it difficult to guarantee the safety and effectiveness of the model in real-world vehicle operation. Summary of the Invention

[0004] In view of the above, it is necessary to propose a verification method, device, controller, vehicle and medium for autonomous driving models to solve the technical problem that existing testing schemes have safety hazards, cannot meet actual testing needs, and make it difficult to guarantee the safety and effectiveness of the models in real vehicle operation.

[0005] Firstly, this application provides a method for verifying an autonomous driving model. The method includes: during vehicle operation, enabling an autonomous driving model to be verified deployed in the vehicle to run in shadow mode, synchronously collecting first environmental data of the vehicle and the driver's human decision-making and planning results, and obtaining, based on the first environmental data and the human decision-making and planning results, risk scenarios where the autonomous driving model to be verified has decision-making biases, and environmental data segments in the first environmental data and human decision-making and planning segments in the human decision-making and planning results corresponding to the risk scenarios; using the human decision-making and planning segments as supervision labels for the environmental data segments, and incrementally training the autonomous driving model to be verified based on the environmental data segments and the supervision labels to obtain an incrementally trained autonomous driving model; controlling the vehicle to run on an open road, and collecting the... The second environmental data of the vehicle is input into the incrementally trained autonomous driving model to obtain the decision planning result of the first model. The second environmental data is then input into a pre-trained scene recognition model to obtain multiple driving scenarios of the vehicle running on open roads. Based on the decision planning result of the first model, the first feature parameters of the incrementally trained autonomous driving model in each driving scenario are statistically analyzed. Based on the human decision planning result, the second feature parameters of the driver in each driving scenario are statistically analyzed. Based on the comparison result of the first feature parameters and the corresponding second feature parameters in each driving scenario, the comprehensive credibility index of the incrementally trained autonomous driving model is calculated. If the comprehensive credibility index is greater than a preset credibility threshold, the incrementally trained autonomous driving model is determined to have passed the verification.

[0006] In the verification method of the autonomous driving model in this application embodiment, the vehicle environment data and the driver's human decision-making and planning results are collected synchronously using shadow mode. This automatically identifies risk scenarios with decision-making biases and uses the human decision-making and planning results as supervisory labels for targeted incremental training of the model. This effectively solves the problems of high cost and incomplete risk scenario coverage associated with traditional methods that rely on manual annotation, thus improving the efficiency and relevance of risk scenario coverage. Based on this, the incrementally trained model is deployed for real-world driving on open roads. Driving scenarios are dynamically classified using a scene recognition model. The characteristic parameters of model decisions and driver decisions under different driving scenarios are statistically analyzed. A comprehensive credibility index is calculated through deviation comparison, and the model's validity is determined based on the comparison between the comprehensive credibility index and a preset threshold. Therefore, this application constructs a multi-level, closed-loop verification system of "automatic risk scenario screening—incremental model training—statistical evaluation on open roads," achieving a quantitative assessment of the uncertainty of autonomous driving model decisions. This significantly improves the generalization ability and credibility of the verification results, while ensuring the safety and effectiveness of real-world testing.

[0007] In some embodiments of this application, before acquiring the risk scenario and the environmental data fragments of the vehicle corresponding to the risk scenario and the model decision planning fragment of the driver, the method further includes: deploying the autonomous driving model to be verified on the vehicle; if a preset edge scenario category is detected during the vehicle's operation, acquiring third environmental data corresponding to the edge scenario category; inputting the third environmental data into the autonomous driving model to be verified to obtain model output data under the edge scenario category; if the model output data does not meet the preset output requirements corresponding to the edge scenario category, training the autonomous driving model to be verified based on the third environmental data corresponding to the edge scenario category and the preset output requirements, until the model output data of the autonomous driving model to be verified under the edge scenario category meets the preset output requirements.

[0008] In some embodiments of this application, the step of obtaining risk scenarios where the autonomous driving model to be verified has decision-making bias based on the first environmental data and the human decision-making planning result includes: inputting the first environmental data into the autonomous driving model to be verified to obtain the second model decision-making planning result, and inputting the first environmental data into the scene recognition model to obtain multiple driving scenarios of the vehicle running in shadow mode, wherein each driving scenario corresponds to an environmental data segment in the first environmental data, a model decision-making planning segment in the second model decision-making planning result, and a human decision-making planning segment in the human decision-making planning result within a time interval; calculating the consistency index between the model decision-making planning segment and the human decision-making planning segment corresponding to each driving scenario; and if the consistency index corresponding to any driving scenario is less than a preset consistency threshold, the any driving scenario is determined as the risk scenario.

[0009] In some embodiments of this application, calculating the consistency index between the model decision planning segment and the human decision planning segment corresponding to each driving scenario includes: obtaining the position control error between the model decision planning segment and the human decision planning segment, and determining a first difference item based on the position control error; obtaining the vehicle speed control deviation between the model decision planning segment and the human decision planning segment, and determining a second difference item based on the vehicle speed control deviation; obtaining the consistency of the decision results between the model decision planning segment and the human decision planning segment, and determining a third difference item based on the consistency of the decision results; and calculating the consistency index corresponding to each driving scenario based on the first difference item, the second difference item, and the third difference item corresponding to each driving scenario.

[0010] In some embodiments of this application, calculating the comprehensive credibility index of the incrementally trained autonomous driving model based on the comparison results of the first feature parameter and the corresponding second feature parameter in each driving scenario includes: comparing the deviations of the first feature parameter and the second feature parameter to obtain a deviation comparison result, wherein the deviation comparison result includes an extreme value deviation comparison result, a central trend deviation comparison result, and a dispersion deviation comparison result; calculating the individual credibility index of each driving scenario based on the deviation comparison result corresponding to each driving scenario; and weighting the individual credibility index of each driving scenario according to a preset scenario weight to obtain the comprehensive credibility index.

[0011] In some embodiments of this application, the step of incrementally training the autonomous driving model to be verified based on the environmental data fragment and the supervision label to obtain the incrementally trained autonomous driving model includes: determining the environmental data fragment and the supervision label as real risk scenario data; generating simulated risk scenario data based on the real risk scenario data, and performing closed-loop simulation verification on the autonomous driving model to be verified based on the simulated risk scenario data; responding to the autonomous driving model to be verified encountering a failure risk scenario in the closed-loop simulation verification, performing randomization augmentation processing on the simulated risk scenario data based on the failure risk scenario to obtain randomized augmented simulated risk scenario data; inputting the randomized augmented simulated risk scenario data into a preset world model to generate virtual risk scenario data, and merging the real risk scenario data and the virtual risk scenario data into an incremental training dataset; and using the incremental training dataset to incrementally train the autonomous driving model to be verified to obtain the incrementally trained autonomous driving model.

[0012] Secondly, this application also provides a verification device for an autonomous driving model, the device comprising: a data acquisition module, configured to, during vehicle operation, enable the autonomous driving model to be verified deployed in the vehicle to run in shadow mode, synchronously collect first environmental data of the vehicle and the driver's human decision-making and planning results, and based on the first environmental data and the human decision-making and planning results, obtain risk scenarios where the autonomous driving model to be verified has decision-making biases, and environmental data segments in the first environmental data and human decision-making and planning segments in the human decision-making and planning results corresponding to the risk scenarios; an incremental training module, configured to use the human decision-making and planning segments as supervision labels for the environmental data segments, and perform incremental training on the autonomous driving model to be verified based on the environmental data segments and the supervision labels to obtain an incrementally trained autonomous driving model; and a road testing module, configured to control the vehicle to run on open roads and collect data. The system uses the second environmental data of the vehicle as input to the incrementally trained autonomous driving model to obtain the decision planning result of the first model, and then inputs the second environmental data into a pre-trained scene recognition model to obtain multiple driving scenarios of the vehicle running on open roads; a feature statistics module is used to count the first feature parameters of the incrementally trained autonomous driving model in each driving scenario based on the decision planning result of the first model, and to count the second feature parameters of the driver in each driving scenario based on the human decision planning result; an index calculation module is used to calculate the comprehensive credibility index of the incrementally trained autonomous driving model based on the comparison result of the first feature parameters and the corresponding second feature parameters in each driving scenario; and a model verification module is used to determine that the incrementally trained autonomous driving model has passed verification if the comprehensive credibility index is greater than a preset credibility threshold.

[0013] Thirdly, this application also provides a controller, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the verification method of the autonomous driving model described in the above embodiments.

[0014] Fourthly, this application also provides a vehicle, the vehicle including the controller described in the above embodiments, the controller being used to execute the verification method of the autonomous driving model described in the above embodiments.

[0015] Fifthly, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the verification method for the autonomous driving model described in the above embodiments.

[0016] Understandably, the verification device for the autonomous driving model of the second aspect, the controller of the third aspect, the vehicle of the fourth aspect, and the computer-readable storage medium of the fifth aspect all correspond to the verification method for the autonomous driving model of the first aspect. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding verification method for the autonomous driving model provided above, and will not be repeated here. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of a verification method for an autonomous driving model provided in an embodiment of this application.

[0018] Figure 2 yes Figure 1 A detailed flowchart of step S100 in the verification method of the autonomous driving model shown.

[0019] Figure 3 yes Figure 1 The diagram shows a detailed flowchart of step S200 in the verification method of the autonomous driving model.

[0020] Figure 4 yes Figure 1 The diagram shows a detailed flowchart of step S500 in the verification method of the autonomous driving model.

[0021] Figure 5 yes Figure 2 The diagram shows a detailed flowchart of step S102 in the verification method of the autonomous driving model.

[0022] Figure 6 This is a schematic diagram of the overall process of a verification method for an autonomous driving model provided in another embodiment of this application.

[0023] Figure 7 This is a schematic diagram of the functional modules of an autonomous driving model verification device provided in an embodiment of this application.

[0024] Figure 8 This is a schematic diagram of the hardware structure of a vehicle provided in one embodiment of this application.

[0025] Explanation of main component symbols Vehicle 1 Controller 10 Memory 11 Processor 12 100 verification devices for autonomous driving models Data acquisition module 110 Incremental Training Module 120 Road test module 130 Feature Statistics Module 140 Indicator Calculation Module 150 Model Validation Module 160 The following detailed description, in conjunction with the accompanying drawings, will further illustrate this application. Detailed Implementation

[0026] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] To provide a clearer understanding of the embodiments of the present invention, the invention will be described in detail below with reference to the accompanying drawings and specific examples. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0028] Please see Figure 1 This is a flowchart illustrating a verification method for an autonomous driving model provided in an embodiment of this application.

[0029] The verification method for the autonomous driving model in this application can be applied to one or more controllers (e.g., Figure 8 The controller 10 shown is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0030] In some embodiments of this application, the controller may be a vehicle (e.g., Figure 8 The onboard equipment of the vehicle shown in 1) includes, for example, driver assistance systems, onboard controllers, and vehicle computers.

[0031] In other embodiments, the controller may also be any electronic product that is communicatively connected to the vehicle, such as a personal computer, tablet computer, etc.

[0032] Specifically, the verification method for autonomous driving models includes the following steps. Depending on different needs, the order of some steps in the flowchart can be changed, and some steps can be omitted.

[0033] Step S100: During vehicle operation, the autonomous driving model to be verified deployed in the vehicle is run in shadow mode, and the first environmental data of the vehicle and the human decision-making and planning results of the driver are collected synchronously. Based on the first environmental data and the human decision-making and planning results, the risk scenarios in which the autonomous driving model to be verified has decision-making biases are obtained, as well as the environmental data segments in the first environmental data and the human decision-making and planning segments in the human decision-making and planning results corresponding to the risk scenarios.

[0034] The shadow mode, to be verified, refers to a mode where the autonomous driving model runs silently in the background of the vehicle but does not actually control the vehicle; the vehicle is driven by the driver. In this mode, the controller synchronously acquires sensor data, map data, and positioning data from the vehicle as primary environmental data, and driver operation data as the result of human decision-making and planning.

[0035] The first environmental data consists of continuous time-series data, including but not limited to sensor data (e.g., visual images, LiDAR point cloud data, vehicle motion data such as wheel speed, acceleration, yaw rate, etc.), map data (e.g., lane geometry and type (solid / dashed lines) in high-precision maps, location and semantic meaning of traffic signs, road slope and curvature, three-dimensional coordinates of guardrails and curbs, etc.), and positioning data (e.g., longitude, latitude, heading angle, pitch angle, roll angle, etc.). Human decision-making and planning results characterize the control strategies adopted by the driver in the current driving scenario, including but not limited to maintaining speed, accelerating, decelerating, stopping, lane keeping, changing lanes, pulling over, detouring, turning left, turning right, and making a U-turn.

[0036] The autonomous driving model to be validated is an end-to-end autonomous driving model based on deep learning. The input is the vehicle's environmental data, and the output is decision control commands. The autonomous driving model is pre-trained offline and deployed in the vehicle's backend in shadow mode for real-time inference but does not participate in vehicle control. The inference results of the autonomous driving model are compared with the driver's human decision planning results to identify risky scenarios with decision bias.

[0037] The risk scenarios are not a pre-defined, fixed list, but are dynamically acquired based on the initial environmental data and human decision-making and planning results. Examples include scenarios like "ghost pedestrians" (sudden pedestrians appearing from behind obstacles) and interactions with vulnerable road users (VRUs), thus improving the relevance and effectiveness of the test data. Each risk scenario corresponds to a segment of environmental data from the initial environmental data and a segment of human decision-making and planning results within a specific time interval.

[0038] In some embodiments of this application, when obtaining risk scenarios where the autonomous driving model to be verified has decision-making bias based on the first environmental data and the human decision-making planning results, the first environmental data is input into the autonomous driving model to be verified to obtain the second model decision-making planning results, and the first environmental data is input into the scene recognition model to obtain multiple driving scenarios of the vehicle running in shadow mode. Each driving scenario corresponds to an environmental data segment in the first environmental data, a model decision-making planning segment in the second model decision-making planning results, and a human decision-making planning segment in the human decision-making planning results within a time interval. The consistency index between the model decision-making planning segment and the human decision-making planning segment corresponding to each driving scenario is calculated. If the consistency index corresponding to any driving scenario is less than a preset consistency threshold, any driving scenario is identified as a risk scenario.

[0039] It should be noted that the risk scenarios regarding decision-making bias in obtaining the autonomous driving model to be validated based on initial environmental data and human decision-making planning results will be discussed later. Figure 2 The embodiments shown are described in detail, and will not be repeated here to avoid repetition.

[0040] Based on this, this application uses the consistency index to quantify the difference between the model's decision and the driver's decision, thereby achieving automatic identification and screening of risk scenarios.

[0041] Step S200: Use human decision-making and planning segments as supervision labels for environmental data segments. Based on the environmental data segments and supervision labels, perform incremental training on the autonomous driving model to be validated to obtain the autonomous driving model after incremental training.

[0042] Specifically, after identifying risk scenarios, this application does not directly use rule-based algorithms for correction, but instead employs a data-driven incremental training approach. Since driver actions typically represent safe and reasonable driving strategies, using human decision-making and planning segments as supervisory labels can guide the autonomous driving model to learn the driver's handling methods in risky scenarios. For example, in a "ghost pedestrian" scenario, the autonomous driving model might predict continuing, while the driver slows down to avoid the obstacle. In this case, the human decision-making and planning segment (the slow-down and avoidance maneuver) is assigned a high weight as a supervisory label.

[0043] In some embodiments of this application, when incrementally training the autonomous driving model to be verified based on environmental data fragments and supervision labels to obtain the incrementally trained autonomous driving model, the environmental data fragments and supervision labels are determined as real risk scenario data; simulated risk scenario data is generated based on the real risk scenario data, and closed-loop simulation verification is performed on the autonomous driving model to be verified based on the simulated risk scenario data; in response to the autonomous driving model to be verified encountering a failure risk scenario in the closed-loop simulation verification, the simulated risk scenario data is randomized and augmented based on the failure risk scenario to obtain randomized augmented simulated risk scenario data; the randomized augmented simulated risk scenario data is input into a preset world model to generate virtual risk scenario data, and the real risk scenario data and virtual risk scenario data are merged into an incremental training dataset; the incremental training dataset is used to incrementally train the autonomous driving model to be verified to obtain the incrementally trained autonomous driving model.

[0044] It should be noted that the process of incrementally training the autonomous driving model to be validated based on environmental data fragments and supervision labels, and obtaining the incrementally trained autonomous driving model, will be discussed in subsequent... Figure 3 The embodiments shown are described in detail, and will not be repeated here to avoid repetition.

[0045] Based on this, this application enables the autonomous driving model to quickly correct decision-making biases in specific scenarios by performing targeted incremental training on the autonomous driving model, thereby improving the decision-making ability of the autonomous driving model in risky scenarios and obtaining an autonomous driving model with better performance after incremental training.

[0046] Step S300: Control the vehicle to run on open roads, collect the vehicle's second environmental data, input the second environmental data into the incrementally trained autonomous driving model to obtain the decision planning result of the first model, and input the second environmental data into the pre-trained scene recognition model to obtain multiple driving scenarios of the vehicle running on open roads.

[0047] In some embodiments of this application, the controller pre-defines the performance standards, test boundaries and verification objectives that the autonomous driving model must meet in open road testing based on the preset functional specifications and design domain (ODD) range. At the same time, it formulates a real vehicle test verification strategy in combination with the diverse requirements of driving scenarios and the requirements for scenario parameter distribution coverage to ensure that the test scenarios cover various typical scenarios and edge scenarios within the ODD range, and that the scenario parameters are evenly distributed.

[0048] Furthermore, after completing incremental training, the incrementally trained autonomous driving model needs to be deployed on vehicles for open road testing and verification. At this point, the vehicle no longer operates in shadow mode; instead, the incrementally trained autonomous driving model takes over vehicle control, realizing the transformation of autonomous driving model verification from background data collection to front-end actual control. Following the real-vehicle testing and verification strategy, the vehicle is controlled to operate on open roads, and second-environment data is collected in real time.

[0049] The second environmental data here differs significantly in timing from the first environmental data in step S100: the first environmental data corresponds to the shadow mode stage before incremental training of the autonomous driving model, while the second environmental data should be in the real vehicle testing stage after incremental training of the autonomous driving model. The second environmental data also includes, but is not limited to, visual images, LiDAR point cloud data, motion state data, etc.

[0050] Furthermore, the collected second environmental data is synchronously input into the incrementally trained autonomous driving model and the pre-trained scene recognition model. The incrementally trained autonomous driving model performs real-time inference on the second environmental data and outputs the first model decision planning results. The first model decision planning results represent the control strategies that the incrementally trained autonomous driving model intends to adopt in the current driving scenario, such as maintaining speed, accelerating, decelerating, stopping, lane keeping, changing lanes, parking on the side of the road, detouring, turning left, turning right, and making a U-turn. At the same time, the scene recognition model performs real-time analysis on the second environmental data, dividing the continuous open road driving process into multiple specific driving scenarios.

[0051] The scene recognition model can be an image classification model based on convolutional neural networks or a temporal classification model based on a deep learning architecture; this embodiment does not impose any restrictions on this. The driving scenarios output by the scene recognition model include, but are not limited to, lane centering, cut-in, cut-out, following control, stop-and-go, starting, lane merging, lane disappearance, lever lane change, efficiency lane change, evasive lane change, navigation lane change, merging, merging out, ramp passage, roundabout, construction zone, VRU interaction, intersection, traffic lights, left turn, right turn, U-turn, narrow road passage, and unexpected pedestrian appearances.

[0052] Based on this, by introducing a scene recognition model, subsequent feature parameter statistics can be performed based on the specific driving context, avoiding misjudgments caused by differences in feature parameters under different driving scenarios.

[0053] In other embodiments, the controller also utilizes the specific driving scenarios segmented by the scene recognition model, combined with the decision planning results of the first model and the corresponding confidence parameters, to conduct model training data verification and model uncertainty assessment based on the set verification objectives, thereby avoiding misjudgments caused by differences in feature parameters under different driving scenarios. Specific verification and assessment methods include, but are not limited to: (1) If the decision planning result of the first model is consistent with bad driving habits in real life under the current driving scenario, it means that there is dirt in the incremental training data corresponding to the current driving scenario. It is necessary to clean and remove the dirty data in the incremental training dataset, reconstruct the incremental training dataset, and perform secondary fine-tuning on the autonomous driving model after incremental training.

[0054] (2) If it is detected that the confidence level of the decision planning result of the first model is not high in the current driving scenario, and the vehicle's driving behavior is conservative and the decision is hesitant, it indicates that the incremental training data corresponding to the current driving scenario is insufficient and lacks diversity. Step S200 needs to be repeated to collect targeted data and perform incremental training for the current driving scenario, and update the autonomous driving model. (3) If the decision deviation of the autonomous driving model is detected to be large in some driving scenarios, and it is found that the training data is leaked or not fully covered, then the effective data in some driving scenarios is supplemented, the training dataset is optimized, and the autonomous driving model is fine-tuned again to ensure that the autonomous driving model is adapted to the complex scenarios of open roads.

[0055] Based on this, after the above verification and evaluation, if the autonomous driving model meets the functional specifications and ODD scope requirements in all test scenarios, and the training data is free from contamination, data leakage, comprehensive coverage and diversity, and the model decision confidence meets the standard and there are no uncertainty anomalies, then the autonomous driving model is determined to have passed the open road test verification; otherwise, if the autonomous driving model fails the open road test verification, the above verification, optimization and retesting steps are repeated until the autonomous driving model meets all verification requirements and the final update of the autonomous driving model is completed.

[0056] Step S400: Based on the decision planning results of the first model, calculate the first feature parameters of the autonomous driving model after incremental training in each driving scenario, and based on the human decision planning results, calculate the second feature parameters of the driver in each driving scenario.

[0057] Specifically, to quantitatively evaluate the performance of the autonomous driving model, this application employs statistical methods to extract features from the testing process. For each identified driving scenario, the controller statistically analyzes the first feature parameters of the first model decision-making and planning results output by the autonomous driving model. Simultaneously, it retrieves the human decision-making and planning results collected in the shadow mode in step S100 and statistically analyzes the second feature parameters of the driver in the same driving scenario.

[0058] Among them, feature parameters refer to statistical quantities used to characterize the quality and safety of driving behavior. The distribution of feature parameters of interest varies in different driving scenarios. For example, in the lane centering driving scenario, the first / second feature parameters include, but are not limited to, vehicle speed, road curvature radius, and vehicle lateral control deviation; in the following control scenario, the first / second feature parameters include, but are not limited to, headway, extreme values ​​and variances of longitudinal acceleration, and braking response delay time; in the lane changing scenario, the first / second feature parameters include, but are not limited to, lane changing execution time, steering angle change rate, and safe distance margin between the vehicle and the vehicles in front and behind in the target lane; in the left turn at an intersection scenario, the first / second feature parameters include, but are not limited to, the average speed when passing through the intersection, the deviation between the steering trajectory and the ideal trajectory, and the timing of avoiding oncoming vehicles.

[0059] Based on this, feature parameters of model decision-making and human decision-making are extracted for each scenario category to provide a quantitative basis for subsequent credibility comparison and evaluation.

[0060] Step S500: Calculate the comprehensive credibility index of the autonomous driving model after incremental training based on the comparison results of the first feature parameter and the corresponding second feature parameter in each driving scenario.

[0061] Among them, the comprehensive credibility index is the core index proposed in this application for quantifying the uncertainty of autonomous driving models.

[0062] In some embodiments of this application, when calculating the comprehensive credibility index of the incrementally trained autonomous driving model based on the comparison results of the first feature parameter and the corresponding second feature parameter under each driving scenario, the first feature parameter and the second feature parameter are compared for deviation to obtain the deviation comparison results. The deviation comparison results include extreme value deviation comparison results, central trend deviation comparison results, and dispersion deviation comparison results. Based on the deviation comparison results corresponding to each driving scenario, the individual credibility index of each driving scenario is calculated. The individual credibility index of each driving scenario is weighted and averaged according to the preset scenario weights to obtain the comprehensive credibility index.

[0063] It should be noted that the calculation of the comprehensive credibility index of the incrementally trained autonomous driving model based on the comparison results of the first feature parameter and the corresponding second feature parameter in each driving scenario will be discussed in subsequent sections. Figure 4 The embodiments shown are described in detail, and will not be repeated here to avoid repetition.

[0064] Based on this, this application constructs a comprehensive credibility index by calculating the degree of deviation between the first feature parameter and the second feature parameter. This index not only considers performance in a single scenario but also integrates the weights of multiple scenario categories, thus comprehensively reflecting the credibility of the autonomous driving model in the overall driving task. For example, higher weights are assigned to safety-critical scenarios (such as pedestrian crossings), making the performance of the autonomous driving model in these scenarios have a greater impact on the comprehensive credibility index.

[0065] Step S600: Determine whether the comprehensive credibility index is greater than the preset credibility threshold.

[0066] The preset credibility threshold is a critical value pre-set according to security specifications and functional requirements. It can be set according to actual performance needs, and this application does not impose any restrictions on it.

[0067] In some embodiments of this application, if the comprehensive confidence index is greater than the preset confidence threshold, it indicates that the model decision of the autonomous driving model after incremental training is close to the human decision of the driver. It is determined that this version of the autonomous driving model has reached the expected performance standard, and the controller continues to execute step S700.

[0068] In some embodiments of this application, if the comprehensive credibility index is less than or equal to the preset credibility threshold, it indicates that the model decision of the autonomous driving model after incremental training deviates significantly from the human decision of the driver, which may pose a safety hazard and still needs to be optimized. The retraining process can be triggered, for example, by supplementing training samples for scenarios with insufficient data, repeating the above incremental training and verification process, i.e., returning to step S200 until the performance of the autonomous driving model meets the standard.

[0069] Step S700: Determine if the incrementally trained autonomous driving model passes validation.

[0070] Specifically, when the overall credibility index is greater than the preset credibility threshold, it indicates that this version of the autonomous driving model has reached the expected performance standard, has sufficient safety and reliability, and can output a verification pass instruction to guide the subsequent freezing and release or large-scale deployment of the autonomous driving model.

[0071] Based on this, through this closed-loop verification process, this embodiment realizes continuous monitoring and iterative optimization of the performance of the autonomous driving model, effectively ensuring the safety of the autonomous driving model operating on open roads.

[0072] In the verification method of the autonomous driving model in this application embodiment, the vehicle environment data and the driver's human decision-making and planning results are collected synchronously using shadow mode. This automatically identifies risk scenarios with decision-making biases and uses the human decision-making and planning results as supervisory labels for targeted incremental training of the autonomous driving model. This effectively solves the problems of high cost and incomplete risk scenario coverage associated with traditional methods that rely on manual annotation, thus improving the efficiency and relevance of risk scenario coverage. Based on this, the incrementally trained autonomous driving model is deployed for real-world driving on open roads. Driving scenarios are dynamically classified using a scene recognition model. The characteristic parameters of model decisions and driver decisions under different driving scenarios are statistically analyzed. A comprehensive credibility index is calculated through deviation comparison, and the autonomous driving model is judged to have passed verification based on the comparison between the comprehensive credibility index and a preset threshold. Based on this, this application constructs a multi-level, closed-loop verification system of "automatic risk scenario screening—incremental model training—statistical evaluation on open roads," achieving a quantitative assessment of the uncertainty of autonomous driving model decisions, significantly improving the generalization ability and credibility of verification results, while ensuring the safety and effectiveness of real-world testing.

[0073] Please see Figure 2 This is a detailed flowchart of step S100 in the verification method of an autonomous driving model provided in an embodiment of this application.

[0074] This embodiment is a detailed description of the foregoing embodiment, further illustrating how to obtain risk scenarios where the autonomous driving model to be verified has decision-making biases based on the first environmental data and human decision-making and planning results. Specifically, it includes the following steps: Step S101: Input the first environmental data into the autonomous driving model to be verified to obtain the decision planning result of the second model, and input the first environmental data into the scene recognition model to obtain multiple driving scenarios of the vehicle running in shadow mode.

[0075] The second model decision planning result represents the decision control commands output by the autonomous driving model to be verified based on the first environmental data, such as maintaining speed, accelerating, decelerating, stopping, lane keeping, changing lanes, parking on the side of the road, detouring, turning left, turning right, and making a U-turn. The second model decision planning result differs significantly in timing from the first model decision planning result in step S300: the first model decision planning result corresponds to the real-vehicle testing phase after incremental model training, while the second model decision planning result corresponds to the shadow mode phase before incremental training of the autonomous driving model.

[0076] The scene recognition model can employ either an image classification model based on convolutional neural networks or a temporal classification model based on a deep learning architecture; this embodiment does not impose any limitations on either. The driving scenarios output by the scene recognition model include, but are not limited to, lane centering, cut-in, cut-out, following control, stop-and-go, starting, lane merging, lane disappearance, lever lane change, efficiency lane change, evasive lane change, navigation lane change, merging, merging out, ramp passage, roundabouts, construction zones, VRU interaction, intersections, traffic lights, left turn, right turn, U-turn, narrow road passage, and unexpected pedestrian appearances.

[0077] Therefore, by introducing a scene recognition model, subsequent consistency assessments can be based on specific driving contexts, avoiding misjudgments caused by differences in control strategies across different driving scenarios. For example, in following-car scenarios, the driver brakes more frequently, while in highway cruising scenarios, the braking frequency is lower. If scene classification is not performed and global data is directly compared, the assessment results will be distorted.

[0078] Step S102: Calculate the consistency index between the model decision planning segment and the human decision planning segment corresponding to each driving scenario.

[0079] Each driving scenario corresponds to an environmental data segment in the first environmental data, a model decision planning segment in the second model decision planning result, and a human decision planning segment in the human decision planning result within a time interval.

[0080] In some embodiments of this application, the consistency index is a core indicator for quantifying the similarity between model decisions and human decisions. For accurate calculation of the consistency index, please refer to [link to relevant documentation]. Figure 5 This application calculates the consistency index between the model decision-making and planning segments and the human decision-making and planning segments for each driving scenario by constructing multi-dimensional difference terms, specifically including: Step S1021: Obtain the position control error between the model decision planning segment and the human decision planning segment, and determine the first difference term based on the position control error.

[0081] Position control error reflects the deviation between the autonomous driving model and the driver in lateral vehicle control. For example, in a lane centering scenario, position control error can be the lateral distance deviation between the planned trajectory of the autonomous driving model and the driver's actual driving trajectory.

[0082] Step S1022: Obtain the speed control deviation between the model decision planning segment and the human decision planning segment, and determine the second difference term based on the speed control deviation.

[0083] Among them, speed control deviation reflects the difference between the autonomous driving model and the driver in longitudinal control. For example, in a following scenario, speed control deviation can be the difference between the autonomous driving model's suggested speed and the driver's actual speed.

[0084] Step S1023: Obtain the consistency of decision results between the model decision planning segment and the human decision planning segment, and determine the third difference term based on the consistency of decision results.

[0085] The consistency of decision-making results reflects the differences in decision-making logic between the autonomous driving model and the driver. The decision space of the model's decision-making planning segment and the human's decision-making planning segment can include, but is not limited to, {maintaining speed, accelerating, decelerating, stopping, lane keeping, changing lanes, parking on the side of the road, detouring, turning left, turning right, and making a U-turn}.

[0086] Specifically, the third difference term is used to quantify this decision-making discrepancy. If the decision action output by the autonomous driving model is consistent with the driver's decision action, the third difference term takes the value of 1 (or a preset high value); otherwise, it takes the value of 0 (or a preset low value). For example, in a scenario where a pedestrian suddenly appears, if the driver's decision is "change lanes" while the autonomous driving model's decision is "stop," the decision results are inconsistent, and the third difference term is 0.

[0087] Based on this, consistency assessment can penetrate the underlying control signals and reach the decision-making logic level, effectively identifying potential scenarios where the control quantities are similar but the decision-making logic is incorrect.

[0088] Step S2024: Calculate the consistency index for each driving scenario based on the first difference item, the second difference item, and the third difference item corresponding to each driving scenario.

[0089] As a concrete example of calculation, the controller calculates the consistency index using a weighted summation method. For instance, the consistency index can be calculated using the following formula: in, It is a consistency index; The weighting coefficient of the first difference term, The weighting coefficients for the second difference term, The weighting coefficient for the third difference term, and The specific values ​​can be dynamically adjusted according to the safety requirements of different driving scenarios, but must meet the following requirements: ; This represents the absolute value of the position control error. This represents the maximum value of the position control error; The control speed output by the autonomous driving model. To help the driver control the vehicle speed, This represents the maximum value of the vehicle speed control deviation. This refers to the decision-making deviation between the autonomous driving model and the driver, i.e., the third difference term.

[0090] It should be noted that the above formulas are merely illustrative and do not constitute a limitation on this application. In other embodiments, more difference terms (such as acceleration deviation, steering angle deviation, etc.) may be introduced, and a nonlinear weighting method may be used to calculate the consistency index.

[0091] Based on this, this application constructs a rigorous screening logic through a multi-dimensional weighted calculation mechanism, which effectively prevents misjudgment of risk scenarios based on a single indicator and ensures that the screened risk scenarios do indeed have cognitive biases.

[0092] Step S103: If the consistency index corresponding to any driving scenario is less than the preset consistency threshold, then any driving scenario is identified as a risk scenario.

[0093] The preset consistency threshold is the baseline for determining whether the autonomous driving model has "learned" the current driving scenario. It can be set according to safety level requirements, for example, to 0.8.

[0094] Specifically, when the consistency index is lower than the consistency threshold, it means that the autonomous driving model to be verified has a significant cognitive bias between its decision-making in the current driving scenario and the driver's decision-making. The autonomous driving model may have cognitive defects or insufficient training data. The controller then marks the current driving scenario as a risky driving scenario and directly triggers the subsequent incremental training process, thereby achieving precise targeting of the performance shortcomings of the autonomous driving model.

[0095] In summary, this application quantifies the difference between the autonomous driving model to be verified and the driver's decision-making through a consistency index, thereby achieving automatic identification and accurate screening of risk scenarios and providing accurate data support for subsequent incremental training.

[0096] Please see Figure 3 This is a detailed flowchart of step S200 in the verification method of an autonomous driving model provided in an embodiment of this application.

[0097] This embodiment is a detailed description of the foregoing embodiments, further illustrating how to incrementally train the autonomous driving model to be validated based on environmental data fragments and supervision labels, thereby obtaining the incrementally trained autonomous driving model. Specifically, it includes the following steps: Step S201: Identify the environmental data fragments and monitoring labels as real-world risk scenario data.

[0098] The real-world risk scenario data refers to raw environmental data fragments collected in shadow mode, filtered by a consistency index to confirm the existence of decision-making biases, and corresponding human decision-making and planning fragments. This data includes real-world sensor noise, ambient lighting changes, and the actual operational logic of human drivers, serving as the foundational dataset for incremental training. For example, if the selected risk scenario is a "ghost peek" (a sudden, unexpected movement in front of a vehicle), then the real-world risk scenario data includes environmental data fragments (e.g., point cloud data, image data) within a specific time interval corresponding to the "ghost peek" driving scenario, as well as supervised labels such as driver decision-making and planning fragments (e.g., steering angle, speed control).

[0099] Step S202: Generate simulated risk scenario data based on real risk scenario data, and perform closed-loop simulation verification on the autonomous driving model to be verified based on the simulated risk scenario data.

[0100] In some embodiments of this application, when generating simulated risk scenario data based on real risk scenario data and performing closed-loop simulation verification on the autonomous driving model to be verified based on the simulated risk scenario data, the real risk scenario data is input into a preset scenario reconstruction model, and the geometric structure, texture, lighting of static participants and the motion trajectory of dynamic participants are extracted from the real risk scenario data to reconstruct a high-fidelity three-dimensional simulation scene, thereby generating simulated risk scenario data.

[0101] The scene reconstruction model can employ Neural Radiation Field (NeRF) technology. NeRF trains a multilayer perceptron (MLP) network to learn the color and density information of spatial points in a scene, enabling the rendering of high-fidelity scene images from any viewpoint. Through NeRF technology, real-world risk scene data collected in the real world can be digitally reconstructed in a virtual environment to obtain simulated scene data, thereby reproducing the physical laws and traffic flow conditions of the real world.

[0102] It should be noted that, in addition to NeRF technology, scene reconstruction models can also employ scene vectorization technology. This involves extracting vector information of static elements (such as lane lines and traffic signs) and dynamic elements (such as vehicles and pedestrians) in the scene using computer vision algorithms and then reconstructing them in a simulation environment. This reconstruction method not only preserves the characteristics of the real scene but also gives the scene editability, providing a foundation for subsequent generalization.

[0103] After reconstructing the simulation scene, closed-loop simulation verification is performed. This includes using the autonomous driving model to be verified as the controller in the simulation environment, granting it complete control over the simulated vehicle. The input to the autonomous driving model is real-time sensor data generated in the simulation environment, and the output is the vehicle's decision-making and control commands. Upon receiving the decision-making and control commands, the simulation engine updates the simulated vehicle's pose, velocity, and acceleration based on the vehicle dynamics model, and simultaneously updates the motion states of surrounding traffic participants (e.g., vehicles, pedestrians, and non-motorized vehicles). Based on the updated scene state, the simulation engine regenerates the sensor data for the next moment and feeds this data back to the autonomous driving model to be verified, forming a closed-loop interactive process of "perception → decision → control → environmental feedback → new perception."

[0104] During the closed-loop simulation, the system continuously monitors whether the simulated vehicle experiences collisions (including collisions with static obstacles and dynamic participants) and whether it violates traffic rules (such as running red lights, driving against traffic, crossing solid lines, and failing to yield). If no of the above failure risk scenarios occur during the entire simulation, the autonomous driving model to be verified is deemed to have passed the verification under the current risk scenario; if any failure risk scenario occurs, the autonomous driving model is deemed to have failed the test under the current risk scenario, triggering the subsequent incremental training process.

[0105] Based on this, this application, through the aforementioned closed-loop simulation verification, can efficiently evaluate the autonomous driving model's ability to cope with specific risk scenarios in a low-cost simulation environment, avoiding the safety risks and resource consumption caused by directly testing dangerous scenarios on real vehicles.

[0106] Step S203: In response to the failure risk scenario occurring in the closed-loop simulation verification of the autonomous driving model to be verified, randomization augmentation processing is performed on the simulation risk scenario data based on the failure risk scenario to obtain the randomized augmented simulation risk scenario data.

[0107] Specifically, for failure risk scenarios, original environmental data and corresponding expected outputs in a real-vehicle environment are re-acquired, and a high-fidelity simulation scenario is reconstructed using a scenario reconstruction model. Based on this, the following parameters in the reconstructed high-fidelity simulation scenario are randomized: Physical parameter randomization: including vehicle dynamics parameters (e.g., tire friction coefficient, suspension stiffness, braking response delay, mass distribution) and sensor noise models (e.g., camera Gaussian noise, lidar point cloud frame drop rate, millimeter-wave radar false alarm rate). Scene variants with physical diversity are generated through random sampling within different ranges. Traffic participant behavior randomization: The initial position, speed, acceleration curve, reaction time, and decision-making strategies (e.g., sudden lane change, emergency braking, crossing, etc.) of dynamic participants such as vehicles and pedestrians are parametrically and randomly adjusted to generate various behavior patterns.

[0108] Based on this, a new set of simulation parameter combinations is generated each time, and each combination corresponds to a randomized augmented simulation risk scenario data. The above randomization process is repeated many times to obtain a large number of simulation scenario datasets of the same type but with different details.

[0109] Step S204: Input the randomized augmented simulated risk scenario data into the preset world model to generate virtual risk scenario data, and merge the real risk scenario data and the virtual risk scenario data into an incremental training dataset.

[0110] Among them, the world model is a scenario generation and evolution model (such as an architecture based on Transformer or diffusion model) that learns to predict the state and dynamic changes of the world. It can generate a large amount of virtual risk scenario data with similar risk characteristics to real risk scenario data but with different details by adjusting scenario parameters (such as lighting, weather, obstacle position, traffic participant behavior, etc.).

[0111] Specifically, the world model takes randomized augmented simulation scenario data (including initial scenario state, physical parameters, and behavioral parameters) as input and outputs a series of virtual risk scenario data. For a real-world "ghost pedestrian" risk scenario (i.e., a pedestrian suddenly crossing the road from behind a parked vehicle), the world model can generalize to multiple virtual risk scenarios with different types of obstructions, different pedestrian appearance times, different pedestrian movement speeds, and different relative distances between vehicles and pedestrians. These virtual scenarios retain the original risk characteristics (e.g., angle of entry, relative speed, collision risk) while having a wider range of parameter distributions (e.g., different road topologies, lighting conditions, weather conditions, and degree of occlusion), thus significantly expanding the coverage of risk scenarios.

[0112] Furthermore, by merging real-world risk scenario data with virtual risk scenario data at a certain ratio (e.g., 1:3 to 1:10), the problem of scarcity of high-risk scenario data in the real world is solved, forming a larger and more diverse incremental training dataset. This dataset is then used for incremental training and fine-tuning of the autonomous driving model to improve its generalization ability and robustness in similar risk scenarios.

[0113] Step S205: Use the incremental training dataset to perform incremental training on the autonomous driving model to be validated, and obtain the autonomous driving model after incremental training.

[0114] Specifically, the incremental training dataset is divided into a training set (80%) and a validation set (20%). Each training sample contains an environmental data fragment and its corresponding supervisory label (i.e., a human decision-making and planning fragment). The environmental data fragment is used as input to the autonomous driving model, and the corresponding supervisory label is used as the learning optimization target. The autonomous driving model is incrementally fine-tuned based on the training set, and the performance of the trained autonomous driving model is verified through the validation set to obtain the incrementally trained autonomous driving model.

[0115] In summary, by using virtual-to-real migration and data augmentation techniques, the diversity of training data has been enriched, and the generalization ability of autonomous driving models in risky scenarios has been improved.

[0116] Please see Figure 4 This is a detailed flowchart of step S500 in the verification method of an autonomous driving model provided in an embodiment of this application.

[0117] This embodiment is a detailed description of the foregoing embodiment, further illustrating how to calculate the comprehensive credibility index of the incrementally trained autonomous driving model based on the first feature parameter and the second feature parameter. Specifically, it includes the following steps: Step S501: Compare the deviation between the first feature parameter and the second feature parameter to obtain the deviation comparison result.

[0118] The deviation comparison results include extreme value deviation comparison results, central tendency deviation comparison results, and dispersion deviation comparison results. Extreme value deviation comparison is used to evaluate the performance boundaries of the autonomous driving model under extreme conditions, such as comparing the maximum lateral deviation of the autonomous driving model in a certain scenario with the maximum lateral deviation of the driver. Central tendency deviation comparison is used to evaluate the average level of the model's behavior, such as comparing the model's average vehicle speed with the driver's average vehicle speed. Dispersion deviation comparison is used to evaluate the stability of the model's behavior, such as comparing the variance of the model's vehicle speed fluctuations with the variance of the driver's vehicle speed fluctuations.

[0119] Based on this, this application can comprehensively characterize the statistical distribution differences between model driving behavior and human driving behavior by comparing these three dimensions.

[0120] Step S502: Calculate a separate credibility index for each driving scenario based on the deviation comparison results corresponding to each driving scenario.

[0121] Among them, the individual credibility index is used to quantify the credibility of the automated driving model in a specific driving scenario.

[0122] As a specific calculation example, a single credibility index can be calculated using the following formula: in, For individual credibility metrics; For driving scenarios The number of feature parameters; For autonomous driving models in driving scenarios Next Statistical values ​​of the first characteristic parameter; For drivers in driving scenarios Next Statistical values ​​of the second characteristic parameter; This represents the minimum value of the first characteristic parameter / the second characteristic parameter. This represents the maximum value of the first feature parameter / the second feature parameter. This represents the mean of the first feature parameter / the second feature parameter. This represents the variance of the first feature parameter / the second feature parameter; The preset weighting coefficients must satisfy the following conditions: .

[0123] Based on this, extreme value deviations are processed by range normalization, and central trend deviations and dispersion deviations are processed by relative error, thus integrating the deviations of the three dimensions into a comprehensive score. The closer the value is to 1, the better the autonomous driving model performs in the current driving scenario. The closer the behavior distribution is to that of the driver, the higher the credibility.

[0124] Step S503: Calculate the weighted average of the individual credibility indexes for each driving scenario according to the preset scenario weights to obtain the comprehensive credibility index.

[0125] Different driving scenarios have varying degrees of impact on safety; for example, the risk weight of cut-in scenarios is usually higher than that of lane centering scenarios.

[0126] As a specific calculation example, the control applies a separate credibility index to all driving scenarios based on preset scenario weights. A weighted average is used to calculate the overall credibility index. The formula for calculating the overall credibility index is as follows: in, To comprehensively assess credibility, This represents the total number of driving scenarios. Preset scene weights for different driving scenarios.

[0127] Based on this, this application uses a weighted average to comprehensively evaluate the overall performance of autonomous driving models in various driving scenarios, avoiding the one-sidedness of evaluation in a single driving scenario.

[0128] In summary, this application, through an evaluation method based on statistical distribution characteristics, can effectively capture the uncertainty of end-to-end autonomous driving model output, providing a scientific and quantitative basis for the safe release of autonomous driving models.

[0129] Please see Figure 6 This is a flowchart illustrating a verification method for an autonomous driving model provided in another embodiment of this application.

[0130] In practical applications, autonomous driving models to be validated are typically trained offline, and the training environment differs from the actual physical, electrical, and network environment of the vehicle. For example, physical differences such as glass transmittance, slight changes in sensor mounting posture due to vehicle load, and the influence of the actual electrical and network environment on communication and inference speeds can all lead to discrepancies between the model input and expectations. Therefore, after deploying the autonomous driving model to be validated in a vehicle, it is necessary to ensure that the autonomous driving model has basic environmental adaptability and to eliminate interference from the vehicle's own physical, electrical, and network environment on the autonomous driving model's inference.

[0131] In view of this, this application is Figure 1 Based on the illustrated embodiment, the autonomous driving model is further validated for environmental adaptability, specifically including the following steps: Step S10: Deploy the autonomous driving model to be validated in the vehicle.

[0132] Among them, the autonomous driving model to be verified refers to the autonomous driving model that has been pre-trained offline or iteratively updated.

[0133] Specifically, the autonomous driving model to be validated is loaded and integrated into the vehicle's controller, enabling the autonomous driving model to receive environmental data in real time, perform forward reasoning, and output decisions.

[0134] Step S11: If a preset edge scene category is detected during vehicle operation, obtain the third environment data corresponding to the edge scene category.

[0135] Among them, the edge scenario category is pre-set based on typical complex driving situations. It refers to special scenarios that have a low probability of occurrence but have a great impact on driving safety, such as extremely small obstacle scenarios when the vehicle speed reaches the functional limit, encounter scenarios of irregular vehicles (such as extra-long trailers and irregular construction vehicles), and sudden changes in road structure in construction areas.

[0136] Specifically, the vehicle operates under the driver's supervision, with real-time monitoring of the driving scenario. When a driving scenario is detected to match a preset edge scenario category, sensor data from the current driving scenario is automatically extracted as third-environment data. It should be noted that the "third-environment data" here differs from the aforementioned "first-environment data" and "second-environment data" in terms of timing; it specifically refers to data fragments used for environmental adaptability verification in the initial stage of autonomous driving model deployment.

[0137] Step S12: Input the third environment data into the autonomous driving model to be verified to obtain the model output data under the edge scene category.

[0138] The model output data includes, but is not limited to, target detection results, trajectory prediction results, and decision planning instructions.

[0139] Step S13: Determine whether the model output data meets the preset output requirements corresponding to the edge scene category.

[0140] The preset output requirements are safety baseline indicators set for different edge scenario categories. For example, for scenarios with extremely small obstacles when the vehicle speed reaches the functional limit, the output requirements may include a detection recall rate of no less than 95%; for scenarios involving encounters with irregularly shaped vehicles, the output requirements may include a classification accuracy rate of no less than 90%; for scenarios involving sudden changes in road structure in construction areas, the output requirements may include path planning that does not intrude into restricted areas and does not result in sudden stops.

[0141] In some embodiments of this application, if the model output data does not meet the preset output requirements corresponding to the edge scene category, it indicates that the autonomous driving model has a cognitive defect in the current edge scene and needs to continue to execute step S14.

[0142] In some embodiments of this application, if the model output data meets the preset output requirements corresponding to the edge scene category, it indicates that the autonomous driving model has the basic ability to handle typical edge scenes, and step S15 is continued.

[0143] Step S14: Train the autonomous driving model to be verified based on the third environment data corresponding to the edge scene category and the preset output requirements.

[0144] Specifically, if the model output data does not meet the preset output requirements corresponding to the edge scene category, the third environment data corresponding to the current edge scene category is used as the supervision label to train the autonomous driving model in a targeted manner until the model output data of the autonomous driving model to be verified meets the preset output requirements under the edge scene category.

[0145] In one example, this training can be incremental fine-tuning. Specifically, the third environment data and supervision labels corresponding to the current edge scene category are used as real edge scene data and input into a preset world model for parameterized generalization to generate virtual edge scene data. The real edge scene data and virtual edge scene data are merged into an incremental training set, and the autonomous driving model is incrementally trained and its parameters are fine-tuned to obtain an updated autonomous driving model. The updated autonomous driving model is then redeployed to the vehicle, and steps S10-S13 are repeated until the output data of the autonomous driving model meets the preset output requirements.

[0146] Based on this, it is ensured that the basic perception and recognition capabilities of the autonomous driving model in the vehicle environment meet the standards, avoiding invalid data collection and model evaluation when basic capabilities are lacking, thereby ensuring the validity of the collected environmental data and avoiding misjudgments caused by environmental incompatibility.

[0147] Step S15: During vehicle operation, the autonomous driving model to be verified deployed in the vehicle is run in shadow mode, and the first environmental data of the vehicle and the human decision-making and planning results of the driver are collected synchronously. Based on the first environmental data and the human decision-making and planning results, the risk scenarios in which the autonomous driving model to be verified has decision-making biases are obtained, as well as the environmental data segments in the first environmental data and the human decision-making and planning segments in the human decision-making and planning results corresponding to the risk scenarios.

[0148] It should be noted that the implementation steps of step S15 and step S100 are the same, and will not be repeated here to avoid repetition.

[0149] Step S16: Use human decision-making and planning segments as supervision labels for environmental data segments. Based on the environmental data segments and supervision labels, perform incremental training on the autonomous driving model to be validated to obtain the autonomous driving model after incremental training.

[0150] It should be noted that the implementation steps of step S16 and step S200 are the same, and will not be repeated here to avoid repetition.

[0151] Step S17: Control the vehicle to run on open roads, collect the second environmental data of the vehicle, input the second environmental data into the incrementally trained autonomous driving model to obtain the decision planning result of the first model, and input the second environmental data into the pre-trained scene recognition model to obtain multiple driving scenarios of the vehicle running on open roads.

[0152] It should be noted that the implementation steps of step S17 and step S300 are the same, and will not be repeated here to avoid repetition.

[0153] Step S18: Based on the decision planning results of the first model, calculate the first feature parameters of the autonomous driving model after incremental training in each driving scenario, and based on the human decision planning results, calculate the second feature parameters of the driver in each driving scenario.

[0154] It should be noted that the implementation steps of step S18 and step S400 are the same, and will not be repeated here to avoid repetition.

[0155] Step S19: Calculate the comprehensive credibility index of the autonomous driving model after incremental training based on the comparison results of the first feature parameter and the corresponding second feature parameter in each driving scenario.

[0156] It should be noted that the implementation steps of step S19 and step S500 are the same, and will not be repeated here to avoid repetition.

[0157] Step S20: Determine whether the comprehensive credibility index is greater than the preset credibility threshold.

[0158] In some embodiments of this application, if the comprehensive confidence index is greater than the preset confidence threshold, it indicates that the model decision of the autonomous driving model after incremental training is close to the human decision of the driver. It is determined that this version of the autonomous driving model has reached the expected performance standard, and the controller continues to execute step S21.

[0159] In some embodiments of this application, if the comprehensive credibility index is less than or equal to the preset credibility threshold, it indicates that the model decision of the autonomous driving model after incremental training deviates significantly from the human decision of the driver, which may pose a safety hazard and still needs to be optimized. The retraining process can be triggered, for example, by supplementing training samples for scenarios with insufficient data, repeating the above incremental training and verification process, i.e., returning to step S16 until the performance of the autonomous driving model meets the standard.

[0160] It should be noted that the implementation steps of step S60 and step S600 are the same, and will not be repeated here to avoid repetition.

[0161] Step S21: Determine if the incrementally trained autonomous driving model passes validation.

[0162] It should be noted that the implementation steps of step S21 and step S700 are the same, and will not be repeated here to avoid repetition.

[0163] In the verification method of the autonomous driving model in this application embodiment, firstly, by deploying the autonomous driving model to be trained to a vehicle and performing edge scene environment verification, the interference of the vehicle's physical (e.g., glass transmittance, sensor installation posture), electrical, and network environment on the model's inference is effectively eliminated, ensuring that the autonomous driving model has basic environmental adaptability, thereby guaranteeing the authenticity and reliability of subsequent risk scene screening and incremental training. Secondly, by using shadow mode to synchronously collect vehicle environmental data and the driver's human decision-making and planning results, risk scenes with decision-making biases are automatically identified, and the human decision-making and planning results are used as supervision labels for targeted incremental training of the autonomous driving model. This effectively solves the problems of high cost and incomplete risk scene coverage of traditional methods that rely on manual annotation, improving the coverage efficiency and targeting of risk scenes. On this basis, the incrementally trained autonomous driving model is deployed on open roads for real-vehicle operation. The driving scenarios are dynamically classified by combining the scene recognition model, and the feature parameters of the model's decisions and the driver's decisions under different driving scenarios are statistically analyzed. A comprehensive credibility index is calculated by comparing the bias, and the autonomous driving model is judged to pass the verification based on the comparison result of the comprehensive credibility index and the preset threshold. Based on this, this application constructs a multi-level, complete closed-loop iterative verification mechanism of "environmental verification → risk screening → incremental training → simulation generalization → open road credibility assessment", which significantly improves the pertinence, efficiency and credibility of autonomous driving model verification while ensuring the safety of real vehicle testing.

[0164] Please see Figure 7 This is a functional module diagram of an autonomous driving model verification device 100 provided in an embodiment of this application.

[0165] In this embodiment, based on the above... Figure 1-6 Using the same concept as the verification method for the autonomous driving model in the illustrated embodiment, this application also provides an autonomous driving model verification device 100, which can be used to execute the above-described autonomous driving model verification method. For ease of explanation, the schematic diagram of the autonomous driving model verification device 100 embodiment only shows the parts related to the embodiments of this application. Those skilled in the art will understand that the illustrated structure does not constitute a limitation on the autonomous driving model verification device 100, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0166] Specifically, the autonomous driving model verification device 100 provided in this application embodiment includes a data acquisition module 110, an incremental training module 120, a road testing module 130, a feature statistics module 140, an index calculation module 150, and a model verification module 160. The data acquisition module 110 is used to enable the autonomous driving model to be verified deployed in the vehicle to run in shadow mode during vehicle operation, synchronously collecting the vehicle's first environmental data and the driver's human decision-making and planning results, and based on the first environmental data and human decision-making and planning results, obtaining risk scenarios where the autonomous driving model to be verified has decision-making biases, and corresponding environmental data fragments in the first environmental data and human decision-making and planning fragments in the human decision-making and planning results; the incremental training module 120 is used to use the human decision-making and planning fragments as supervision labels for the environmental data fragments, and to perform incremental training on the autonomous driving model to be verified based on the environmental data fragments and supervision labels to obtain the incrementally trained autonomous driving model; the road testing module 130 is used to control the vehicle to run on open roads, collect the vehicle's second environmental data, and perform incremental training on the third environmental data fragments and human decision-making and planning results; The second environmental data is input into the incrementally trained autonomous driving model to obtain the decision planning result of the first model, and the second environmental data is input into the pre-trained scene recognition model to obtain multiple driving scenarios of the vehicle running on open roads; the feature statistics module 140 is used to calculate the first feature parameters of the incrementally trained autonomous driving model in each driving scenario based on the decision planning result of the first model, and to calculate the second feature parameters of the driver in each driving scenario based on the human decision planning result; the index calculation module 150 is used to calculate the comprehensive credibility index of the incrementally trained autonomous driving model based on the comparison result of the first feature parameters and the corresponding second feature parameters in each driving scenario; the model verification module 160 is used to determine that the incrementally trained autonomous driving model has passed verification if the comprehensive credibility index is greater than the preset credibility threshold.

[0167] Specific limitations regarding the verification device 100 for the autonomous driving model can be found in the limitations of the verification method for the autonomous driving model described above, and will not be repeated here. Each module in the aforementioned verification device 100 for the autonomous driving model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the controller in hardware form, or stored in the memory of the controller in software form, so that the processor can call and execute the operations corresponding to each module.

[0168] Please see Figure 8 This is a schematic diagram of the hardware structure of the controller 10 provided in an embodiment of this application.

[0169] The controller 10 provided in this application embodiment includes, but is not limited to, a memory 11, a processor 12, and a computer program stored in the memory 11 and executable on the processor 12, such as a damper control program. When the computer program is executed by the processor 12, it implements the verification method of the autonomous driving model as described in the above embodiment.

[0170] Figure 8 Only the controller 10, which includes a memory 11 and a processor 12, is shown. Those skilled in the art will understand that... Figure 8 The structure shown does not constitute a limitation on the controller 10 and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0171] In some embodiments of this application, the controller 10 can be communicatively connected to devices such as desktop computers, laptops, handheld computers, and cloud servers.

[0172] In some embodiments of this application, the controller 10 can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control device.

[0173] In some embodiments of this application, the controller 10 may further include network devices and / or client devices. These network devices include, but are not limited to, a single network server, a server group consisting of multiple network servers, and a cloud server based on cloud computing, consisting of a large number of hosts or network servers.

[0174] In some embodiments of this application, the network where the controller 10 is located includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, virtual private network (VPN), etc.

[0175] In some embodiments of this application, memory 11 stores multiple computer-readable instructions to implement a method for verifying an autonomous driving model, and processor 12 can execute multiple instructions to implement the method for verifying an autonomous driving model as described in the above embodiments.

[0176] Specifically, the processor 12's implementation method for the above instructions can be found in [reference]. Figures 1-6 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0177] Those skilled in the art will understand that the schematic diagram is merely an example of the controller 10 and does not constitute a limitation on the controller 10. The controller 10 can be a bus topology or a star topology. The controller 10 may also include more or fewer other hardware or software than shown in the diagram, or different component arrangements. For example, the controller 10 may also include an input / output controller 10, a network access device, etc.

[0178] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, in... Figure 8 The symbol is represented by only one arrow, but this does not mean that there is only one bus or one type of bus. The bus is configured to implement communication between memory 11 and processor 12, etc.

[0179] It should be noted that controller 10 is only an example. Other existing or future electronic products that are suitable for this application should also be included within the scope of protection of this application and are incorporated herein by reference.

[0180] In some embodiments of this application, the processor 12 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 12 is the control core of the controller 10, connecting various components of the controller 10 via various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., executing a damper control program) and calls data stored in the memory 11 to perform various functions of the controller 10 and process data.

[0181] The processor 12 executes the operating system of the controller 10 and various installed applications. The processor 12 executes these applications to implement the steps in the various embodiments of the verification method for an autonomous driving model described above, for example... Figures 1-6 The steps are shown.

[0182] For example, a computer program can be divided into one or more modules / units, one or more of which are stored in memory 11 and executed by processor 12 to complete this application. One or more modules / units can be a series of computer-readable instruction segments capable of performing a specific function, which describe the execution process of the computer program in controller 10. For example, a computer program can be divided into... Figure 7 The module shown.

[0183] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute a portion of the verification method for an autonomous driving model according to various embodiments of this application.

[0184] If the modules / units integrated into controller 10 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware devices. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above.

[0185] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory, and other types of memory.

[0186] This application also provides a computer-readable storage medium (not shown) storing computer-readable instructions, which are executed by a processor in a controller 10 to implement a verification method for an autonomous driving model according to any of the above embodiments.

[0187] Specifically, computer-readable storage media can be non-volatile or volatile. Computer-readable storage media include flash memory, portable hard drives, multimedia cards, card-type memories (e.g., SD memory, DX memory, etc.), magnetic storage, magnetic disks, optical disks, etc. In some embodiments, memory 11 can be an internal storage unit of the controller 10, such as the portable hard drive of the controller 10. In other embodiments, memory 11 can also be an external storage device of the controller 10, such as a plug-in portable hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the controller 10. Memory 11 can be used not only to store application software and various types of data installed on the controller 10, such as the code of a damper control program, but also to temporarily store data that has been output or will be output.

[0188] Furthermore, the computer-readable storage medium may primarily include a stored program area and a stored data area, wherein the stored program area may store the operating system, applications required for the functions, etc.; and the stored data area may store data created based on the use of blockchain nodes, etc.

[0189] In the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, "multiple" means two or more.

[0190] In the embodiments of this application, it should be noted that, unless otherwise expressly specified and limited, the word "for example" is used to indicate an example, illustration, or description. Any embodiment or design scheme described as "for example" in the embodiments of this application should not be construed as being better or more advantageous than other embodiments or design schemes. Specifically, the use of the word "for example" is intended to present the relevant concepts in a specific manner.

[0191] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0192] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more features.

[0193] In the description of this application, it should be noted that, unless otherwise explicitly stated and limited, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Furthermore, the character " / " in this application generally indicates that the preceding and following related objects have an "or" relationship.

[0194] Unless otherwise specified, all steps in this application may be performed sequentially or randomly. For example, if a method includes steps A and B, it means that the method may include steps A and B performed sequentially, or it may include steps B and A performed sequentially. For example, if a method may also include step C, it means that step C may be added to the method in any order. For example, the method may include steps A, B, and C, or it may include steps A, C, and B, or it may include steps C, A, and B, etc.

[0195] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

[0196] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0197] In the various embodiments of this application, the functional modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0198] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices described in the specification may also be implemented by a single unit or device through software or hardware.

[0199] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application.

Claims

1. A method for verifying an autonomous driving model, characterized in that, The method includes: During vehicle operation, the autonomous driving model to be verified deployed on the vehicle is run in shadow mode, and the first environmental data of the vehicle and the human decision-making and planning results of the driver are collected synchronously. Based on the first environmental data and the human decision-making and planning results, the risk scenarios in which the autonomous driving model to be verified has decision-making biases are obtained, as well as the environmental data segments in the first environmental data and the human decision-making and planning segments in the human decision-making and planning results corresponding to the risk scenarios. The human decision-making and planning segment is used as the supervision label for the environmental data segment. Based on the environmental data segment and the supervision label, the autonomous driving model to be verified is incrementally trained to obtain the autonomous driving model after incremental training. The vehicle is controlled to operate on an open road. Second environmental data of the vehicle is collected, and the second environmental data is input into the incrementally trained autonomous driving model to obtain the decision planning result of the first model. The second environmental data is then input into a pre-trained scene recognition model to obtain multiple driving scenarios of the vehicle operating on an open road. Based on the decision planning results of the first model, the first feature parameters of the incrementally trained autonomous driving model in each driving scenario are statistically analyzed, and based on the human decision planning results, the second feature parameters of the driver in each driving scenario are statistically analyzed. Based on the comparison results of the first feature parameter and the corresponding second feature parameter in each driving scenario, the comprehensive credibility index of the incrementally trained autonomous driving model is calculated. If the comprehensive credibility index is greater than the preset credibility threshold, the incrementally trained autonomous driving model is determined to have passed the verification.

2. The verification method for the autonomous driving model as described in claim 1, characterized in that, Before acquiring the risk scenario and the corresponding vehicle environmental data fragments and driver model decision-making planning fragments, the method further includes: The autonomous driving model to be verified is deployed on the vehicle. If a preset edge scene category is detected during the vehicle's operation, the third environmental data corresponding to the edge scene category is obtained. The third environmental data is input into the autonomous driving model to be verified to obtain the model output data under the edge scene category; If the model output data does not meet the preset output requirements corresponding to the edge scene category, the autonomous driving model to be verified is trained according to the third environment data corresponding to the edge scene category and the preset output requirements until the model output data of the autonomous driving model to be verified under the edge scene category meets the preset output requirements.

3. The verification method for the autonomous driving model as described in claim 1, characterized in that, The risk scenarios where the autonomous driving model to be verified, based on the first environmental data and the human decision-making and planning results, has decision-making biases include: The first environmental data is input into the autonomous driving model to be verified to obtain the second model decision planning result, and the first environmental data is input into the scene recognition model to obtain multiple driving scenarios of the vehicle running in shadow mode. Each driving scenario corresponds to an environmental data segment in the first environmental data, a model decision planning segment in the second model decision planning result, and a human decision planning segment in the human decision planning result within a time interval. Calculate the consistency index between the model decision planning segment and the human decision planning segment for each driving scenario; If the consistency index corresponding to any driving scenario is less than the preset consistency threshold, then the driving scenario is identified as the risk scenario.

4. The verification method for the autonomous driving model as described in claim 3, characterized in that, The calculation of the consistency index between the model decision planning segment and the human decision planning segment corresponding to each driving scenario includes: Obtain the position control error between the model decision planning segment and the human decision planning segment, and determine the first difference term based on the position control error; Obtain the speed control deviation between the model decision planning segment and the human decision planning segment, and determine a second difference term based on the speed control deviation; Obtain the consistency of decision results between the model decision planning segment and the human decision planning segment, and determine a third difference term based on the consistency of the decision results; Based on the first difference item, the second difference item, and the third difference item corresponding to each driving scenario, the consistency index corresponding to each driving scenario is calculated.

5. The verification method for the autonomous driving model as described in claim 1, characterized in that, The step of calculating the comprehensive credibility index of the incrementally trained autonomous driving model based on the comparison results of the first feature parameter and the corresponding second feature parameter in each driving scenario includes: The first feature parameter and the second feature parameter are compared to obtain a deviation comparison result, wherein the deviation comparison result includes an extreme value deviation comparison result, a central trend deviation comparison result, and a dispersion deviation comparison result. Based on the deviation comparison results corresponding to each driving scenario, calculate a separate credibility index for each driving scenario; The individual credibility index for each driving scenario is weighted and averaged according to the preset scenario weights to obtain the comprehensive credibility index.

6. The verification method for the autonomous driving model as described in claim 1, characterized in that, The step of incrementally training the autonomous driving model to be verified based on the environmental data fragment and the supervision label to obtain the incrementally trained autonomous driving model includes: The environmental data fragments and the monitoring labels are identified as real risk scenario data; Based on the real-world risk scenario data, simulated risk scenario data is generated, and the autonomous driving model to be verified is subjected to closed-loop simulation verification based on the simulated risk scenario data. In response to the failure risk scenario occurring in the closed-loop simulation verification of the autonomous driving model to be verified, the simulation risk scenario data is randomized and augmented based on the failure risk scenario to obtain the randomized and augmented simulation risk scenario data. The randomized augmented simulated risk scenario data is input into a preset world model to generate virtual risk scenario data, and the real risk scenario data and the virtual risk scenario data are merged into an incremental training dataset. The autonomous driving model to be verified is incrementally trained using the incremental training dataset to obtain the incrementally trained autonomous driving model.

7. A verification device for an autonomous driving model, characterized in that, The device includes: The data acquisition module is used to enable the autonomous driving model to be verified deployed in the vehicle to run in shadow mode during the vehicle's operation, synchronously collect the vehicle's first environmental data and the driver's human decision-making and planning results, and obtain the risk scenarios where the autonomous driving model to be verified has decision-making biases based on the first environmental data and the human decision-making and planning results, as well as the environmental data segments in the first environmental data and the human decision-making and planning segments in the human decision-making and planning results corresponding to the risk scenarios. The incremental training module is used to use the human decision-making and planning segment as the supervision label of the environmental data segment, and to perform incremental training on the autonomous driving model to be verified based on the environmental data segment and the supervision label to obtain the autonomous driving model after incremental training. The road testing module is used to control the vehicle to run on open roads, collect the second environmental data of the vehicle, input the second environmental data into the incrementally trained autonomous driving model to obtain the decision planning result of the first model, and input the second environmental data into the pre-trained scene recognition model to obtain multiple driving scenarios of the vehicle running on open roads. The feature statistics module is used to count the first feature parameters of the incrementally trained autonomous driving model in each driving scenario based on the decision planning results of the first model, and to count the second feature parameters of the driver in each driving scenario based on the human decision planning results. The index calculation module is used to calculate the comprehensive credibility index of the incrementally trained autonomous driving model based on the comparison results of the first feature parameter and the corresponding second feature parameter under each driving scenario. The model verification module is used to determine that the incrementally trained autonomous driving model has passed verification if the comprehensive credibility index is greater than a preset credibility threshold.

8. A controller, characterized in that, The controller includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the verification method for the autonomous driving model as described in any one of claims 1 to 6.

9. A vehicle, characterized in that, The vehicle includes a controller as described in claim 8, the controller being configured to perform a verification method for an autonomous driving model as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code that can be invoked by a processor to execute the verification method for the autonomous driving model as described in any one of claims 1 to 6.

Images