Method and system for generating man-machine cooperation conflict test case

Abstract

The invention provides a man-machine cooperation conflict test case generation method and system, and the method comprises the following steps: S1, collecting a vehicle motion state, a driving operation behavior and a driver physiological state, carrying out the multi-modal data fusion modeling, constructing a personalized driving style model, and calculating driving style parameters; s2, detecting man-machine cooperation conflict behaviors in real time, and quantifying man-machine cooperation conflict levels; s3, inputting the man-machine collaboration conflict level and the driving style parameter as conditional variables into a generative adversarial imitation learning model to generate a man-machine collaboration conflict test case; s4, verifying the dynamics rationality of the man-machine cooperation conflict test case through a physical engine; and S5, distributing and iterating the man-machine collaborative conflict test case generation model by adopting a federated learning mechanism. The man-machine cooperation conflict test case generated by the method can truly simulate man-machine right and liability games and behavior contradictions existing in the actual driving process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic control of automobiles, in particular to a man-machine collaborative conflict test scene generation method and system. BACKGROUND

[0002] With the development of automatic driving technology, testing and verification has become a key link to ensure its safety and reliability. At present, automatic driving testing mainly relies on two ways of real vehicle testing and virtual simulation testing. Although real vehicle testing can reflect the real road performance, it is high in cost, long in cycle, and limited by environment and regulations, making it difficult to cover a large number of scenes. With the development of automatic driving systems to a higher level, the requirements for testing scale and complexity are rapidly increasing, and the limitations of real vehicle testing are increasingly prominent.

[0003] CN114492157B provides an automatic driving test scene generation method based on individualized driver model, which simulates the stages of human drivers in driving experience based on genetic and evolutionary ideas combined with reinforcement learning, and gradually trains driver models at different growth stages, including novice driver model, experienced driver model and skilled driver model; based on the driver models at different growth stages, according to the different individualized driving characteristics of human drivers, further train individualized driver models, including aggressive driver model, conservative driver model, provocative driver model and cooperative driver model; according to the test requirements, combine the driver models at different growth stages and different individualized driver models according to the corresponding demand ratio to generate the corresponding target test scene.

[0004] CN114862156B provides an emotion-driven individual driver model customized test scene generation method, including: based on the existing emotion model, selecting specific emotions that affect driving vehicle behavior; pre-impose specific emotional influence on the test driver, then collect the corresponding driving data of the test driver on the driving simulator; for the collected driving data, first perform imitation learning, then perform reinforcement learning Q-Learning to obtain generalized driving data of different styles; combine driving data of different styles according to a specific ratio, determine hyperparameters through the fireworks optimization algorithm to generate efficient test scenes for automatic driving cars; determine the verification index of the efficient test scene; select the decision-making method of the system to be tested and verify it through the verification index.

[0005] However, the above method mainly focuses on the test scene of driver proficiency and emotional state, and does not effectively model the dynamic interaction behavior between human drivers and automatic driving systems in typical confrontation situations such as control struggle and instruction conflict. Therefore, the generated test cases are difficult to realistically simulate the human-machine responsibility game and behavior contradiction existing in the actual driving process, limiting their effectiveness in evaluating the safety and coordination of human-machine collaboration. SUMMARY

[0006] The purpose of the present application is to provide a human-machine collaboration conflict test scene generation method and system, which dynamically generates human-machine collaboration conflict test cases based on individualized driving style modeling by taking driving style and human-machine collaboration conflict behavior as conditional variables, and can realistically simulate the human-machine responsibility game and behavior contradiction existing in the actual driving process.

[0007] To achieve the above purpose, the present application provides a human-machine collaboration conflict test case generation method, which includes the following steps: S1: collect vehicle motion state, driving operation behavior and driver physiological state, perform multi-modal data fusion modeling, build an individualized driving style model, and calculate driving style parameters; S2: real-time detect human-machine collaboration conflict behavior and quantify human-machine collaboration conflict level; S3: input the human-machine collaboration conflict level and driving style parameters as conditional variables into the generative adversarial imitation learning model, and dynamically generate human-machine collaboration conflict test cases from the generative adversarial imitation learning model; S4: verify the dynamics of the human-machine collaboration conflict test cases by the physics engine, and eliminate invalid test cases; S5: use the federated learning mechanism to distribute and iterate the human-machine collaboration conflict test case generation model.

[0008] The technical effects achieved after adopting this technical solution are: by fusing vehicle motion state, driving operation behavior and driver physiological state, and performing multi-modal data fusion modeling, the depth quantification and individualized representation of driver style are realized, providing accurate data basis for generating highly simulated test cases; by real-time detecting human-machine collaboration conflict behavior and quantifying human-machine collaboration conflict level, human-machine conflict scenes are introduced into the test cases; then taking driver style and human-machine collaboration conflict behavior as conditional variables, the generative adversarial imitation learning model is used to dynamically generate human-machine collaboration conflict test cases based on individualized driving style modeling, which can realistically simulate the human-machine responsibility game and behavior contradiction existing in the actual driving process, and combined with the physics engine and the federated learning mechanism, it provides strong support for the safety verification of automatic driving systems.

[0009] Preferably, the driving style parameters are calculated by the following weighted fusion formula: K = a1 x K_vehicle + a2 x K_operation + a3 x K_physiological Wherein a1, a2, a3 are weight coefficients, satisfying a1+a2+a3=1, K is the driving style parameter, K_vehicle is the vehicle motion state score, K_operation is the driving operation behavior score, and K_physiological is the driving personnel physiological state score.

[0010] The technical effects achieved after adopting the technical scheme are as follows: the driving style is fused with the data of three dimensions of the vehicle motion state, the driving operation behavior, and the driver physiological state, and the test case generated by weighted summation can not only reflect the long-term behavior characteristics of the driver, but also capture the instantaneous decision changes caused by physiological factors such as fatigue and tension, greatly improving the authenticity and individualization of the human-machine conflict scene simulation; and the adjustability of the weight coefficients a1, a2, and a3 allows the generation strategy to be flexibly customized according to different test targets, thereby ensuring comprehensive coverage of the test case while enhancing the directional generation capability for specific high-risk scenes.

[0011] Preferably, the physiological state of the driving personnel includes at least one of heart rate variability, blink frequency, and skin electrical response; the vehicle motion state includes at least one of following distance, longitudinal acceleration, lateral acceleration, and lane deviation; and the driving operation behavior includes at least one of steering wheel angular velocity, steering wheel torque, accelerator pedal opening degree, and emergency braking frequency.

[0012] The technical effects achieved after adopting the technical scheme are as follows: the specific composition of the multi-modal data is accurately defined, wherein the physiological state includes physiological indicators such as heart rate variability, blink frequency, and skin electrical response; the vehicle motion state includes dynamic parameters such as following distance, longitudinal / lateral acceleration, and lane deviation; and the driving operation behavior includes control characteristics such as steering wheel angular velocity, torque, accelerator pedal opening degree, and emergency braking frequency; the cooperative use of these data can capture the implicit physiological state (such as tension and fatigue) and operation intention of the driver, so that the calculation of the driving style parameter has both physiological response authenticity and behavior representation accuracy, thereby generating conflict scenes that are more in line with the complex decision-making mechanism of humans; and these data ensure the collectability, interpretability, and reproducibility of the model input data, significantly improving the transparency and reliability of the test case generation process, and ultimately providing high-value test cases with both physiological authenticity and engineering practicality for human-machine collaborative driving systems.

[0013] Preferably, the driving style is classified according to the driving style parameter: Conservative type: K<0.3; Steady type: 0.3≤K≤0.7; Aggressive: K > 0.7.

[0014] The technical effects achieved after adopting the technical scheme are: through the explicit classification of the driving style parameters (conservative: K < 0.3; stable: 0.3 ≤ K ≤ 0.7; aggressive: K > 0.7), the standardization and quantifiable classification of the driver behavior characteristics are realized, thereby significantly improving the pertinence and coverage of the test case generation: on the one hand, the classification method can accurately match the typical behavior patterns of drivers with different styles, ensure that the generated conflict test cases can truly restore the cautious hesitation of conservative drivers, the balanced decision-making of stable drivers and the adventurous tendency of aggressive drivers, and greatly enhance the diversity and realistic restoration degree of the human-machine interaction conflict scene.

[0015] Preferably, the human-machine collaborative conflict behavior includes at least one of an angle difference between the driver steering wheel torque and the automatic driving system control instruction, a difference between the driver pedal operation and the automatic driving system control instruction, a difference between the driver's gaze on the road and the target road of the automatic driving system control instruction, a change in the driver's heart rate variability, and a change in the driver's skin galvanic response.

[0016] The technical effects achieved after adopting the technical scheme are: through the explicit definition of the multi-dimensional performance of the human-machine collaborative conflict behavior including the operation layer (angle difference of steering wheel torque, difference of pedal operation), the cognitive layer (gaze target deviation) and the physiological layer (change in heart rate variability, change in skin galvanic response), the multi-modal deep fusion perception and quantification of the conflict behavior are realized, through the introduction of the driver's gaze target deviation and real-time physiological stress response (such as heart rate variability reduction representing tension, skin galvanic response enhancement representing alertness), the implicit disagreement between the driver and the automatic driving system in the intention understanding and emotional response can be captured earlier and more accurately, thereby realizing the early prediction and graded warning of potential conflicts. This conflict behavior definition based on multi-modal fusion provides extremely rich and close-to-real human reaction feature inputs for the generation of the adversarial learning model, so that it can generate high-fidelity, full-dimensional human-machine conflict test cases not only in operation opposition, but also containing cognitive imbalance and physiological stress, thereby completely improving the test coverage and verification depth of the automatic driving system on the complexity of human-machine co-driving.

[0017] Preferably, the human-machine collaborative conflict level is calculated by the following formula: Conflict_Level = β1 x Δθ norm + β2 x Pedal_Conflict norm + β3 x Eye_deviation norm + β4 x (1 / HRV norm )+ β5 x GSR_mean norm ; wherein Conflict_Level is the human-robot collaboration conflict level, wherein β1, β2, β3, β4, β5 are weight coefficients, satisfying β1+β2+β3+β4+β5=1, Δθ norm is the normalized steering wheel operation angle difference, Pedal_Conflict norm is the normalized pedal operation difference indicator, Eye_deviation norm is the normalized gaze road deviation, HRV norm is the normalized heart rate variability, GSR_mean norm is the normalized mean skin galvanic response.

[0018] The technical effects achieved after adopting the technical scheme are as follows: by introducing the calculation formula of the human-robot collaboration conflict level, the fine and multi-dimensional objective evaluation of the conflict intensity is realized: firstly, the calculation formula comprehensively integrates the external behavior conflict (steering wheel angle difference, pedal operation difference), cognitive attention distraction (gaze road deviation), and internal physiological stress response (reduced heart rate variability and enhanced skin galvanic response), so that the conflict level determination is no longer limited to a single operation difference, but deeply integrates the physiological state of the driver, greatly improving the accuracy and robustness of the evaluation; secondly, through normalization processing and flexible configuration of the weight coefficients (β1 to β5), the dimensional differences of the multi-source data are eliminated, and the evaluation emphasis can be dynamically adjusted according to the conflict type (such as control right contention and attention conflict), so that different conflict intensities from slight intervention to serious confrontation can be accurately distinguished; finally, this quantitative conflict level provides stable and reliable condition input for the generative adversarial imitation learning model, ensuring that the generated test cases can accurately cover different tension levels of human-robot conflict scenarios.

[0019] Preferably, the human-robot collaboration conflict level is classified according to the human-robot collaboration conflict level: Low conflict: Conflict_Level<0.4; Medium conflict: 0.4≤Conflict_Level<0.7 High conflict: Conflict_Level≥0.7.

[0020] The technical effects achieved after adopting the technical scheme are: by quantifying the human-machine collaborative conflict level into three clear levels, namely, low, medium and high, the standardization and fine hierarchical management of the conflict intensity are realized, wherein the grading method converts the continuous comprehensive conflict metric value into discrete semantic labels, so that the generation of test cases and the response strategy of the autonomous driving system can realize accurate matching and differentiated processing (for example, only recording logs for low conflict, initiating warnings for medium conflict, and triggering emergency takeover for high conflict), greatly enhancing the pertinence and effectiveness of the test; and the grading mechanism provides clear conditional constraints for the generation of the adversarial imitation learning model, enabling it to generate conflict scenarios covering the full spectrum from minor differences to extreme conflicts, thereby systematically verifying the human-machine interaction reliability and safety of the autonomous driving system under different human-machine conflict levels.

[0021] Preferably, the adversarial training of the adversarial imitation learning model comprises: the generator receives random noise and the conditional variable, and outputs synthetic scene data; the discriminator receives real scene data or the synthetic scene data and the conditional variable, and outputs a scalar probability value; the discriminator is updated by fixing the generator, and the generator is updated by fixing the discriminator, and the adversarial training is performed until Nash equilibrium is reached.

[0022] The technical effects achieved after adopting the technical scheme are: the adversarial training mechanism forces the generator to learn the underlying distribution of real human-machine conflict scenarios under the strict constraints of specific conflict levels and driving styles, thereby generating high-quality synthetic data that conforms to physical laws and is rich in details, effectively solving the problems of high cost of real conflict data acquisition and scarcity of long-tail scenarios; the discriminator ensures the high consistency of the generated scene and the target condition by jointly discriminating the matching of data and conditions (such as whether the generated aggressive driving style scene corresponds to a high conflict level); through the dynamic optimization process of alternating training until Nash equilibrium, the generator and the discriminator continuously evolve in the game, and the generated data approaches or even surpasses the real data distribution in terms of realism and diversity, thereby providing the autonomous driving system with a test scenario library that covers a wide range, has a reasonable distribution, and contains a large number of extreme cases.

[0023] Preferably, the federated learning mechanism further comprises: a central server initializes a human-machine collaborative conflict test case generation model and distributes it to each node; each node trains and individualizes the human-machine collaborative conflict test case generation model distributed to each node based on local data; each node encrypts and uploads the model parameters of the human-machine collaborative conflict test case generation model to the central server for secure aggregation; the human-machine collaborative conflict test case generation model is updated and distributed to each node for the next round of iteration.

[0024] The technical effects achieved after the technical scheme are as follows: the mechanism utilizes the generalization ability and robustness of the data value promotion model of all nodes, can strictly protect data privacy, can accurately focus on core test targets such as individual driving styles and man-machine conflicts, and finally efficiently generates test cases that can expose system weaknesses, thereby providing strong power for safety verification of the automatic driving system.

[0025] The application further provides a man-machine collaborative conflict test scene generation system for implementing any of the above technical solutions, which comprises a central server and a plurality of distributed nodes in communication connection with the central server, wherein the distributed nodes are arranged in vehicle terminals or simulators; the central server runs a man-machine collaborative conflict test case generation model; the distributed nodes run a local man-machine collaborative conflict test case generation model; the man-machine collaborative conflict test case generation model and the local man-machine collaborative conflict test case generation model both comprise the following functional modules: a driving style modeling model for calculating driving style parameters; a man-machine conflict detection model for real-time detection and calculation of the level of man-machine collaborative conflict; a generative adversarial imitation learning model for generating man-machine collaborative conflict test cases; and a verification model for dynamic reasonableness verification of the man-machine collaborative conflict test cases.

[0026] The technical effects achieved after the technical scheme are as follows: the system arranges the distributed nodes in the vehicle terminals or simulators, directly collects multi-modal data locally, wherein the data collection module is directly embedded in the terminal to realize millisecond-level real-time data acquisition, the driving style modeling model and the man-machine conflict detection model are used to calculate the driving style parameters and the conflict level in real time, the response speed of the system is greatly improved, the generative adversarial imitation learning model is used to generate individualized test cases according to the driving style parameters and the conflict level, the verification module is used to perform dynamic reasonableness verification on the test cases generated by integrating the physical engine, invalid test cases are automatically removed, and the quality of the generated test cases is ensured. Therefore, the system can generate realistic, diverse and high-risk man-machine conflict test scenes on a large scale by using an efficient, self-learning and reasonable automatic method. BRIEF DESCRIPTION OF DRAWINGS

[0027] The above and / or additional aspects and advantages of the application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which: Figure 1 A man-machine collaborative conflict test case generation method step flowchart is provided for the embodiments of the application. DETAILED DESCRIPTION

[0028] In order to make the above objectives, characteristics and advantages of the present application more apparent, clear and comprehensible, the technical solutions in the embodiments of the present application will be described below clearly and completely. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0029] The embodiment of the present application provides a human-machine collaborative conflict test case generation method, and the generation method comprises the following steps: S1: collecting vehicle motion state, driving operation behavior and driving personnel physiological state, performing multi-modal data fusion modeling, constructing a personalized driving style model, and calculating driving style parameters; S2: real-time detection of human-machine collaborative conflict behavior, and quantification of human-machine collaborative conflict level; S3: inputting the human-machine collaborative conflict level and the driving style parameters as condition variables into the generative adversarial imitation learning model, and dynamically generating human-machine collaborative conflict test cases by the generative adversarial imitation learning model; S4: verifying the dynamics rationality of the human-machine collaborative conflict test cases by a physical engine, and automatically eliminating invalid test cases; and S5: distributing and iterating the human-machine collaborative conflict test case generation model by using a federated learning mechanism. For details, see Figure 1 as shown.

[0030] The present scheme realizes deep quantification and personalized representation of the driving style by fusing the vehicle motion state, driving operation behavior and driving personnel physiological state, and performing multi-modal data fusion modeling, provides a precise data basis for generating highly realistic test cases, introduces human-machine conflict scenarios in the test cases by real-time detection of human-machine collaborative conflict behavior and quantification of human-machine collaborative conflict level, and dynamically generates human-machine collaborative conflict test cases based on personalized driving style modeling by the generative adversarial imitation learning model with the driving style and human-machine collaborative conflict behavior as condition variables, which can truly reflect the human-machine right and responsibility game and behavior contradiction existing in the actual driving process, and solves the problems of lack of human-machine interaction modeling, insufficient scene reality and data island in the traditional test method by combining the physical engine and the federated learning mechanism. The core breakthrough is to include the driving personnel psychological motivation and physiological indicators into the conflict decision, and to generate a high-fidelity test scene through an interpretable mathematical mapping, thereby providing a more reliable verification framework for the automatic driving system.

[0031] In some embodiments of the present application, the driving style parameters are calculated by the following weighted fusion formula: K=α1×K_vehicle+α2×K_opersation+α3×K_physiological wherein a1, a2, a3 are weight coefficients, satisfying a1+a2+a3=1, K is a driving style parameter, K_vehicle is a vehicle motion state score, K_operation is a driving operation behavior score, and K_physiological is a driver physiological state score.

[0032] Specifically, the driving style comprehensive evaluation index driving style parameter K is obtained by weighted summation of the vehicle motion state score K_vehicle, the driving operation behavior score K_operation, and the driver physiological state score K_physiological. The core idea of the vehicle motion state score is to quantify the motion intensity and stability of the vehicle itself, and to reflect the direct influence of the driving behavior on the dynamic characteristics of the vehicle. Aggressive driving behavior is usually manifested as greater longitudinal and lateral acceleration, and a higher score indicates that the vehicle motion is more intense and unstable. The core idea of the driving operation behavior score is that frequent steering wheel fine tuning, large angle steering, and frequent hard braking are typical manifestations of aggressive driving or distraction, and a higher score indicates that the driving operation is more aggressive and unstable. The core idea of the driver physiological state score is that the decrease of heart rate variability (HRV), the increase of galvanic skin response (GSR), or the change of eye movement pattern such as gaze and blink frequency change can objectively expose the increase of cognitive load and fatigue, which makes monitoring these indicators more accurate in assessing the suitable driving state of the driver. A higher value usually indicates that the physiological state of the driver is worse, indicating that the driver is more nervous, fatigued, or distracted. That is, the same driving behavior may have relatively stable physiological indicators (proficiency, confidence) for some drivers, while others may have extremely high galvanic skin response (nervousness, risk-taking). Therefore, physiological data help to distinguish different internal motivations and states under similar external behaviors.

[0033] The driving style of the present scheme fuses data in three dimensions of vehicle motion state, driving operation behavior, and driver physiological state, and obtains the test case by weighted summation, so that the generated test case not only reflects the long-term behavior characteristics of the driver, but also captures the instantaneous decision changes caused by physiological factors such as fatigue and nervousness, greatly improving the authenticity and individualization of human-machine conflict scene simulation. Secondly, the adjustability of the weight coefficients a1, a2, and a3 allows flexible customization of the generation strategy according to different test targets, thereby ensuring comprehensive coverage of the test case while enhancing the directional generation capability for specific high-risk scenarios.

[0034] In some embodiments of the present application, the physiological state of the driver includes at least one of heart rate variability, blink frequency, galvanic skin response; the vehicle motion state includes at least one of following distance, longitudinal acceleration, lateral acceleration, lane deviation; the driving operation behavior includes at least one of steering wheel angular velocity, steering wheel torque, accelerator pedal opening, emergency braking frequency.

[0035] Specifically, the heart rate variability (HRV) can be obtained by a steering wheel capacitive sensor, a smart watch / wearable device or a seat built-in sensor; the blink frequency can be obtained by a non-contact identification through a cabin camera combined with a computer vision algorithm; the galvanic skin response (GSR) can be obtained by a special bracelet or a contact electrode on the steering wheel; the longitudinal / lateral acceleration can be obtained by calculating data from the vehicle's inertial measurement unit or wheel speed sensor; the lane deviation can be obtained by a lane recognition system of a vehicle-mounted camera or a high-precision positioning module (such as RTK-GPS); the following distance can be obtained by emitting a laser beam and measuring the reflection time or identifying the front vehicle from the point cloud; the steering wheel angular velocity, the steering wheel torque, the accelerator pedal opening and the emergency braking frequency can be obtained by reading from the vehicle CAN bus.

[0036] According to the above data, the calculation formula of the vehicle motion state score K_vehicle is: K_vehicle=a 1× long_std+a2×lat_std+a3×lane_deviation_std+a4×K_d; Wherein long_std is the normalized longitudinal acceleration standard deviation, reflecting the degree of acceleration / deceleration; lat_std is the normalized lateral acceleration standard deviation, reflecting the degree of steering; lane_deviation_std is the normalized standard deviation of the distance between the vehicle mass center and the lane center line, reflecting the stability of the lane keeping ability; K_d is the following distance score; a1, a2, a3, a4 are weight coefficients, and satisfy a1+a2+a3+a4=1.

[0037] Also, when K_d=0 (when THW≥T_safe, the following distance is very safe); K_d=1 (when THW≤T_danger, the following distance is extremely dangerous); K_d changes linearly between T_danger and T_safe. Wherein THW is the headway time, the time required for two vehicles to reach each other's position at the current speed; T_safe is the safety time threshold, T_danger is the danger time threshold; T_safe can be set to more than 3 seconds (such as 3.5 seconds), T_danger can be set to 1 second or less (such as 0.8 seconds).

[0038] The calculation formula of the driving operation behavior score K operation is: K_opersation = b1 x steering_std + b2 x brake_event_rate + b3 x throttle_std; steering_std is a normalized steering wheel angle speed standard deviation, reflecting the smoothness or urgency of the steering operation; brake_event_rate is a normalized emergency braking event frequency, reflecting the aggressiveness and safety awareness of the driving; throttle_std is a normalized accelerator pedal opening standard deviation, reflecting the smoothness of the throttle operation, and the higher the value, the more "deep one foot shallow one foot" throttle operation; b1, b2, and b3 are weight coefficients, and satisfy b1 + b2 + b3 = 1.

[0039] The calculation formula of the driver physiological state score K physiological is: K_physiological = c1 x (1 / HRV_std) + c2 x GSR_mean + c3 x blink_rate; HRV_std is a normalized heart rate variability standard deviation, and a decrease in HRV is usually related to stress and fatigue, and a decrease in HRV indicates a poor state, so the higher the value of 1 / HRV_std, the worse the state of the driver; GSR_mean is a normalized skin galvanic response average value, and an increase in GSR usually indicates emotional arousal, tension or stress; blink_rate_norm is the blink rate, and an abnormal increase or decrease in blink rate may be related to fatigue or high concentration; c1, c2, and c3 are weight coefficients, and satisfy c1 + c2 + c3 = 1.

[0040] The present scheme precisely defines the specific composition of multi-modal data, wherein the driver physiological state includes physiological indicators such as heart rate variability, blink rate, and skin galvanic response; the vehicle motion state includes dynamic parameters such as following distance, longitudinal / lateral acceleration, and lane deviation; and the driving operation behavior includes steering wheel angle speed, torque, accelerator pedal opening, and emergency braking frequency. The cooperative use of these data can capture the implicit physiological state (such as tension and fatigue) of the driver and the operation intention, so that the calculation of the driving style parameter has both physiological response authenticity and behavior representation accuracy, thereby generating a conflict scene that is more in line with the complex decision-making mechanism of humans; and these data ensure the collectability, interpretability and reproducibility of the model input data, significantly improving the transparency and reliability of the test case generation process, and finally providing a high-value test case for the human-machine collaborative driving system, which has both physiological authenticity and engineering practicality.

[0041] In some embodiments of the present application, the driving style is classified according to the driving style parameter (K): Conservative: K < 0.3; Steady: 0.3 < K < 0.7; Aggressive: K > 0.7.

[0042] When the driving style is conservative, the vehicle speed is stable, the acceleration and deceleration are gentle, the following distance is far, and the lane changing and overtaking frequency is extremely low; when the driving style is steady, the driving behavior is close to the average level, and there is no obvious risk tendency; when the driving style is aggressive, the vehicle speed is fast, the sudden acceleration and sudden braking are frequent, the following distance is short, and the lane changing and overtaking are active.

[0043] The present scheme realizes the standardization and quantifiable classification of the driving behavior characteristics of the driver by the explicit classification of the driving style parameter (conservative: K < 0.3; steady: 0.3 < K < 0.7; aggressive: K > 0.7), thereby significantly improving the pertinence and coverage of the test case generation: on the one hand, the classification method can accurately match the typical behavior patterns of drivers with different styles, ensuring that the generated conflict test cases can truly restore the cautious hesitation of conservative drivers, the balanced decision-making of steady drivers, and the adventurous tendency of aggressive drivers, greatly enhancing the diversity and realistic restoration of the human-machine interaction conflict scene.

[0044] In some embodiments of the present application, the human-machine collaborative conflict behavior includes at least one of the angle difference between the driver's steering torque and the automatic driving system control instruction, the difference between the driver's pedal operation and the automatic driving system control instruction, the difference between the driver's gaze on the road and the target road of the automatic driving system control instruction, the change in the driver's heart rate variability, and the change in the driver's skin electric response.

[0045] Wherein, the angle difference Δθ of the driver steering wheel torque and the automatic driving system control instruction is the difference (Δα = θ_human - θ_ads) between the steering wheel torque angle (θ_human) applied by the driver and the target angle (θ_ads) of the automatic driving system (ADS) control instruction, which directly represents the instantaneous struggle strength and intention difference of both parties in the vehicle lateral control; the acceleration / brake pedal operation difference is the direct intention confrontation between the driver and the ADS in the vehicle longitudinal control, specifically, the driver steps on the brake pedal deeply while the ADS system is still executing the acceleration instruction, or the human requests acceleration while the ADS system is executing braking; the gaze road deviation is the offset angle between the driver's gaze focus captured by the eye tracker and the target road planned to be traveled by the ADS system, which deeply reveals the decrease of the driver's trust in the ADS decision or understanding difference; the decrease of HRV is a strong signal that the driver is in a high stress state, which reduces his patience and tolerance and amplifies his dissatisfaction with the system behavior, thereby directly pushing up the human-machine conflict level; the peak frequency and amplitude of GSR are perfect indicators for quantifying instantaneous emotional response, for example, a sudden SCR peak can directly correspond to a specific human-machine conflict event (such as system emergency takeover, false alarm).

[0046] The present scheme realizes multi-modal deep fusion perception and quantification of conflict behaviors by clearly defining the multi-dimensional performance of human-machine collaborative conflict behaviors in the operation layer (steering wheel torque angle difference, pedal operation difference), the cognitive layer (gaze road target deviation) and the physiological layer (heart rate variability change, skin electric response change), and by introducing the driver's gaze target deviation and real-time physiological stress response (such as heart rate variability reduction representing tension and skin electric response enhancement representing alertness), the implicit differences in intention understanding and emotional response between the driver and the automatic driving system can be captured earlier and more accurately, thereby realizing the early prediction and graded warning of potential conflicts. This conflict behavior definition based on multi-modal fusion provides extremely rich and close-to-human-real-response feature inputs for the generative adversarial learning model, so that it can generate high-fidelity, full-dimensional human-machine conflict test cases not only containing operation opposition, but also containing cognitive imbalance and physiological stress, thereby completely improving the test coverage and verification depth of the automatic driving system for human-machine co-driving complexity.

[0047] In some embodiments of the present application, the human-machine collaborative conflict level is calculated by the following formula: Conflict_Level = β1 × Δθ norm + β2 × Pedal_Conflict norm + β3 × Eye_deviation norm + β4 × (1 / HRV norm )+ β5 × GSR_mean norm ; wherein Conflict Level is the human-robot collaboration conflict level, wherein β1, β2, β3, β4, β5 are weight coefficients, satisfying β1+β2+β3+β4+β5=1, Δθ norm is the normalized steering wheel operation angle difference, Pedal_Conflict norm is the normalized pedal operation difference indicator, Eye_deviation norm is the normalized gaze road deviation, HRV norm is the normalized heart rate variability, GSR_mean norm is the normalized skin galvanic response average value.

[0048] The scheme realizes fine and multi-dimensional objective evaluation of the conflict intensity by introducing the calculation formula of the human-robot collaboration conflict level: firstly, the calculation formula comprehensively integrates the external behavior conflict (steering wheel angle difference, pedal operation difference), cognitive attention distraction (gaze road deviation), and internal physiological stress response (reduced heart rate variability and enhanced skin galvanic response), so that the conflict level determination is no longer limited to a single operation difference, but deeply integrates the physiological state of the driver, greatly improving the accuracy and robustness of the evaluation; secondly, through normalization processing and flexible configuration of the weight coefficients (β1-β5), the dimensional differences of multi-source data are eliminated, and the evaluation emphasis can be dynamically adjusted according to the conflict type (such as control right contention and attention conflict), so that different conflict intensities from slight intervention to serious confrontation can be accurately distinguished; finally, the quantitative conflict level provides stable and reliable condition input for the generative adversarial imitation learning model, ensuring that the generated test cases can accurately cover different tension levels of human-robot conflict scenarios.

[0049] In some embodiments of the present application, the conflict level is classified according to the human-robot collaboration conflict level: Low conflict: Conflict Level < 0.4; Medium conflict: 0.4 ≤ Conflict Level < 0.7 High conflict: Conflict Level ≥ 0.7.

[0050] The technical solution quantifies the man-machine conflict level into three clear levels: low (<0.4), medium ([0.4, 0.7)), and high (>=0.7), realizes the standardized and fine-grained hierarchical management of the conflict intensity, and the grading method converts the continuous comprehensive conflict metric value into discrete semantic labels, so that the generation of test cases and the response strategy of the autonomous driving system can realize accurate matching and differentiated processing (for example, only need to record logs for low conflict, initiate warning for medium conflict, and trigger emergency takeover for high conflict), greatly enhancing the pertinence and effectiveness of the test; and the grading mechanism provides clear conditional constraints for the generative adversarial imitation learning model, enabling it to generate scenarios covering all spectrum of conflicts from minor differences to extreme antagonism, thereby systematically verifying the man-machine interaction reliability and safety of the autonomous driving system under different man-machine conflict levels.

[0051] Low conflict: the human-machine intention is basically consistent, the operation is smooth, and it belongs to the normal co-driving state; medium conflict: there is an identifiable operation divergence, the driver may be slightly distracted or dissatisfied, and there is potential risk; high conflict: there is a serious divergence in human-machine intention, danger is imminent or has occurred, and the driver may be in a state of tension, anger or dangerous operation.

[0052] In some embodiments of the present application, the adversarial training of the generative adversarial imitation learning model includes: the generator receives random noise and the conditional variable, and outputs synthetic scene data; the discriminator receives real scene data or the synthetic scene data and the conditional variable, and outputs a scalar probability value representing the confidence of data authenticity; the discriminator is updated by fixing the generator, and the generator is updated by fixing the discriminator, and the discriminator is updated for adversarial training, and the training continues until Nash equilibrium.

[0053] Specifically, the core of the GAIL / CGAIL model is to learn the distribution characteristics of the real data through the adversarial training of the generator and the discriminator. Among them, is to learn the distribution characteristics of the real data through the adversarial training of the generator and the discriminator. The generator (G) receives random noise z and optional condition variable c (driving style parameter K value, human-machine conflict level Conflict_Level value), and outputs synthetic scene data G(z|c); the discriminator (D) receives real scene data (s, a) or generated scene data G(z|c) and the corresponding condition variable c, and outputs a scalar probability value representing the confidence of the data authenticity. Among them, the adversarial training process adopts an alternating optimization strategy. First, fix the generator and update the discriminator. Sample a batch of real scene data (s, a) and its corresponding condition variable c from the real data, and sample a batch of generated scenes G(z|c) from the generator. Update the parameters of D by the gradient ascent method to improve its ability to distinguish real and generated data. Secondly, fix the discriminator and update the generator: sample a batch of random noise z and condition variable c, and update the parameters of G by the policy gradient method (such as REINFORCE) or direct back propagation (when G is a differentiable network) to maximize the confidence of the generated data passing D, logD(G(z|c)). The training continues until the Nash equilibrium is reached or the performance meets the requirements. The Nash equilibrium here refers to the game between the generator and the discriminator reaching a balanced state, and the parameters of the generator and the discriminator tend to be stable, and the training loss no longer changes significantly. When the system reaches the Nash equilibrium, the average output probability of the discriminator to the generated data is close to 0.5, that is, the probability that the discriminator considers the generated data to be real is 50%, which indicates that the generated data and the real data cannot be distinguished. That is, the scalar probability value reaches 50%.

[0054] Furthermore, CGAIL's conditional generation mechanism is its core advantage in generating test cases for autonomous driving. It allows the test generation process to be closely integrated with personalized driving styles and the dynamic quantification of human-machine conflict, achieving precise and controllable scenario generation. In CGAIL, the K-value serves as a conditional variable input to the generator and discriminator, directly influencing the "personality" and behavioral patterns of the generated scenarios. By adjusting the K-value, CGAIL can seamlessly cover a spectrum of driving styles from conservative to aggressive, generating matching test scenarios and achieving a comprehensive assessment of the risks that drivers of different styles may face. When the driving style is conservative, if the conditional variable c contains a low K-value, the generator tends to generate low-risk scenarios, such as a large following distance (THW>2.5s), gentle acceleration changes (longitudinal acceleration <1.5m / s²), early and gradual lane changes (steering angular velocity <10° / s), and conservative responses to surrounding vehicle behavior. When the driving style is aggressive, if the variable c contains a high K value, the generator will generate high-risk scenarios, such as extremely short following distance (THW<1.0s), intense acceleration and braking behavior (longitudinal acceleration>3.0m / s²), frequent sudden lane changes (steering angular velocity>30° / s), and provocative interactive behaviors (such as close-range cutting in and vying for right-of-way).

[0055] The dynamic quantitative metric Conflict_Level reflects the degree of difference between the autonomous driving system and the human driver's control intentions in real time. In CGAIL, the target Conflict_Level can be used as a condition variable to directly guide the generator to create test scenarios with a specific conflict intensity. After receiving the target Conflict_Level through the condition variable, the generator adjusts the scenario parameters to achieve that conflict level. For example, to achieve a high conflict level, the generator may shorten the cutting distance of the vehicle in front, increase the speed of the background vehicle, or create visual obstructions to delay the system's perception, thereby inducing a decision-making divergence between the driver and the system. When generating a low-conflict scenario with Conflict_Level < 0.4, the generator will generate a human-machine conflict... Figure 1 Contextualized scenarios include: smooth acceleration / deceleration consistent with driver expectations, lane-changing maneuvers performed by the system matching driver expectations, and system responses to the environment matching driver judgment. When generating high-conflict scenarios (Conflict_Level > 0.7), the generator creates scenarios with severe discrepancies between human and machine intentions to expose system weaknesses, such as emergency takeover requests where the system fails to respond to danger in time, forcing the driver to intervene urgently (e.g., grabbing the steering wheel, slamming on the brakes); erroneous intervention where overly conservative system decisions cause driver discomfort (e.g., unnecessary emergency braking), leading to conflict; and control struggles where the autonomous driving system and driver operate the vehicle simultaneously and in opposite directions (e.g., the driver brakes while the system accelerates).

[0056] The adversarial training mechanism forces the generator to learn the underlying distribution of real human-machine conflict scenarios under strict constraints of specific conflict levels and driving styles, thereby generating high-quality synthetic data that conforms to physical laws and is rich in details, effectively solving the problem of high cost of real conflict data acquisition and scarcity of long-tail scenarios; the discriminator ensures the high consistency of the generated scene and the target condition by jointly judging the matching of the data and the condition (such as whether the generated aggressive driving style scene corresponds to a high conflict level); through the dynamic optimization process of alternating training until Nash equilibrium, the generator and the discriminator continuously evolve in the game, and the finally generated data approaches or even surpasses the real data distribution in terms of realism and diversity, thereby providing an automatic driving system with a test scene library that covers a wide range, has a reasonable distribution, and contains a large number of extreme cases.

[0057] In some embodiments of the present application, the federated learning mechanism further includes: a central server initializes a human-machine collaborative conflict test case generation model and distributes it to each node; each node trains and individualizes the human-machine collaborative conflict test case generation model distributed to each node based on local data; each node encrypts and uploads the model parameters of the human-machine collaborative conflict test case generation model to the central server for secure aggregation; the human-machine collaborative conflict test case generation model is updated and distributed to each node for the next round of iteration.

[0058] The federal learning is essentially a distributed machine learning technology, which aims to realize the distributed training mode of "data not moving and model moving" on the basis of guaranteeing the data privacy safety and legal compliance. Specifically, a central server initializes a man-machine collaborative conflict test case generation model, wherein the man-machine collaborative conflict test case generation model comprises a driving style modeling model, a man-machine conflict detection model and a generative adversarial imitation learning model, and is distributed to all vehicles or simulation test nodes participating in the federal learning; each vehicle or node uses locally collected driving data to locally train and fine-tune the received global model, and the local training data comprises the man-machine conflict level calculated in real time by the personalized driving style K value and the multi-dimensional information of the vehicle motion state; in order to prevent sensitive information leakage, each vehicle or node only uploads the encrypted update (such as gradient or weight) of the model parameter to the central aggregation server after local training, and the original driving data is always kept locally; after the central server collects the encrypted parameters from multiple nodes, a secure aggregation algorithm (such as FedAvg) is performed to generate an updated global model, and weights are assigned according to the data quality, quantity or importance of each node to avoid the influence of the data deviation of some nodes on the performance of the global model; the updated global model is again distributed to each node, which uses it for the next round of local training, and the cycle is repeated, so that the global model evolves in the process of absorbing the wisdom of all nodes, and does not need to store any original data. The mechanism uses the data value of all nodes to improve the generalization ability and robustness of the model, strictly protects data privacy, accurately focuses on personalized driving style and man-machine conflict and other core test targets, and finally efficiently generates test cases that can expose system weaknesses, providing strong power for the safety verification of the automatic driving system.

[0059] The application further provides a man-machine collaborative conflict test scene generation system for realizing the system of any of the above technical solutions, which comprises a central server and a plurality of distributed nodes in communication connection with the central server; wherein the distributed nodes are deployed in vehicle terminals or simulators; the central server runs a man-machine collaborative conflict test case generation model; the distributed nodes run a local man-machine collaborative conflict test case generation model; the man-machine collaborative conflict test case generation model and the local man-machine collaborative conflict test case generation model both comprise the following functional modules: a driving style modeling model for calculating driving style parameters; a man-machine conflict detection model for real-time detection and calculation of the level of man-machine collaborative conflict; a generative adversarial imitation learning model for generating man-machine collaborative conflict test cases; and a verification model for verifying the dynamics rationality of the man-machine collaborative conflict test cases; the distributed nodes further comprise a data acquisition module for acquiring vehicle motion states, driving operation behaviors and driving personnel physiological states.

[0060] The system directly localizes the collection of multi-modal data by deploying distributed nodes in the vehicle terminal or simulator, wherein the data collection module is directly embedded in the terminal to achieve millisecond-level real-time data acquisition, the driving style modeling model and the human-machine conflict detection model are used to calculate the driving style parameters and the conflict level in real time, which greatly improves the response speed of the system, the generative adversarial imitation learning model generates personalized test cases using the driving style parameters and the conflict level, and the verification module generates test cases for dynamic reasonableness verification by integrating a physical engine, automatically eliminates invalid test cases, and ensures the quality of the generated test cases. Therefore, the system can generate realistic, diverse, and high-risk human-machine conflict test scenarios on a large scale through an efficient, self-learning, and reasonable automated method.

[0061] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for generating human-in-the-loop conflict test cases, the method comprising: The generation method comprises the following steps: S1: Collecting vehicle motion state, driving operation behavior and driving personnel physiological state, performing multi-modal data fusion modeling, constructing a personalized driving style model, and calculating driving style parameters; S2: Real-time detection of human-machine collaborative conflict behavior, and quantification of the human-machine collaborative conflict level; S3: According to the human-machine collaborative conflict level and the driving style parameters as condition variables, inputting into the generative adversarial imitation learning model, and dynamically generating the human-machine collaborative conflict test case by the generative adversarial imitation learning model; S4: Verifying the dynamics rationality of the human-machine collaborative conflict test case by a physical engine, and eliminating invalid test cases; S5: Distributing and iterating the human-machine collaborative conflict test case generation model by using a federated learning mechanism.

2. The generation method of claim 1, wherein, The driving style parameters are calculated by the following weighted fusion formula: K = α1×K_vehicle + α2×K_opersation + α3×K_physiological Wherein α1, α2, α3 are weight coefficients, satisfying α1+α2+α3=1, K is the driving style parameter, K_vehicle is the vehicle motion state score, K_operation is the driving operation behavior score, and K_physiological is the driving personnel physiological state score.

3. The generation method according to claim 2, wherein the driving personnel physiological state comprises at least one of heart rate variability, blink frequency, and skin electrical response; The vehicle motion state comprises at least one of following distance, longitudinal acceleration, lateral acceleration, and lane deviation; The driving operation behavior comprises at least one of steering wheel angular velocity, steering wheel torque, accelerator pedal opening degree, and emergency braking frequency.

4. The generation method of claim 2, wherein, According to the driving style parameters, the driving style is classified as: Conservative type: K<0.3; Steady type: 0.3≤K≤0.7; Aggressive type: K>0.

7.

5. The generation method of claim 1, wherein, The human-machine collaborative conflict behavior comprises at least one of the angle difference between the driver's steering wheel torque and the automatic driving system control instruction, the difference between the driver's pedal operation and the automatic driving system control instruction, the difference between the driver's gaze on the road and the target road of the automatic driving system control instruction, the change of the driver's heart rate variability, and the change of the driver's skin electrical response.

6. The generation method of claim 5, wherein, The human-machine collaborative conflict level is calculated by the following formula: Conflict_Level = β1 x Δθ norm + β2 x Pedal_Conflict norm + β3 x Eye_deviation norm + β4 x (1 / HRV norm ) + β5 x GSR_mean norm ; wherein Conflict Level is the human-robot collaboration conflict level, wherein β1, β2, β3, β4, β5 are weight coefficients, satisfying β1+β2+β3+β4+β5=1, Δθ norm is the normalized steering wheel operation angle difference, Pedal_Conflict norm is the normalized pedal operation difference index, Eye_deviation norm is the normalized gaze road deviation, HRV norm is the normalized heart rate variability, GSR_mean norm is the normalized skin galvanic response average value.

7. The generation method of claim 6, wherein, According to the human-machine collaborative conflict level, the human-machine collaborative conflict level is classified as: Low conflict: Conflict_Level<0.4; Medium conflict: 0.4≤Conflict_Level<0.7 High conflict: Conflict_Level≥0.

7.

8. The generation method of claim 1, wherein, The adversarial training of the generative adversarial imitation learning model comprises: The generator receives random noise and the condition variables, and outputs synthetic scene data; the discriminator receives real scene data or the synthetic scene data and the condition variables, and outputs a scalar probability value; By fixing the generator, updating the discriminator and fixing the discriminator, updating the discriminator for adversarial training, and training until Nash equilibrium is reached.

9. The generation method of claim 1, wherein, The federated learning mechanism comprises the following steps: The central server initializes the human-machine collaborative conflict test case generation model and distributes it to each node; Each node trains and individualizes the human-machine collaborative conflict test case generation model based on local data; Each node encrypts the model parameters of the human-machine collaborative conflict test case generation model and uploads them to the central server for secure aggregation; The human-machine collaborative conflict test case generation model is updated and distributed to each node for the next round of iteration.

10. A human-in-the-loop conflict test scenario generation system for implementing the generation method of any one of claims 1 to 9, characterized in that, The system comprises: A central server and a plurality of distributed nodes connected in communication with the central server; The distributed nodes are deployed in vehicle terminals or simulators; The central server runs the human-machine collaborative conflict test case generation model; The distributed nodes run local human-machine collaborative conflict test case generation models; The human-machine collaborative conflict test case generation model and the local human-machine collaborative conflict test case generation model each comprise the following functional modules: A driving style modeling model for calculating driving style parameters; A human-machine conflict detection model for real-time detection and calculation of the level of human-machine collaborative conflict; A generative adversarial imitation learning model for generating the human-machine collaborative conflict test case; A verification model for verifying the dynamics of the human-machine collaborative conflict test case; The distributed nodes further comprise a data acquisition module for acquiring the vehicle motion state, driving operation behavior and driving personnel physiological state.

Images