Interaction cognitive ability evaluation method for unmanned driving

Abstract

The invention provides an interactive cognitive ability evaluation method for unmanned driving. The method comprises the following steps: constructing a layered scene library; designing a pressure factor, injecting the pressure factor into each scene, and carrying out scene testing; in the scene testing process, testing data are collected, and features are extracted; calculating a cognitive entropy-interaction entropy; and carrying out comprehensive scoring and grade authentication. According to the interactive cognitive ability evaluation method oriented to unmanned driving, the cognitive and interactive process of an unmanned driving system can be efficient and predictable, the uncertainty in the decision-making and interactive process of the unmanned driving system is quantified through entropy in the information theory, a pressure test scene is designed and injected, the change of the entropy value is evaluated, and the evaluation accuracy of the interactive cognitive ability of the unmanned driving system is improved. Therefore, the comprehensive ability score is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned driving, and particularly to an interaction cognitive ability evaluation method for unmanned driving. BACKGROUND

[0002] Unmanned driving and automatic driving describe the external form, while automatic driving and autonomous driving are two completely different technical routes. Automatic driving is essentially teaching a robot to drive through a program by installing various sensors, computer systems and related software, starting a computer program, and letting the car drive automatically. It is a software-defined car. Autonomous driving must have a subject of driving cognition. As an intelligent agent of the driver, autonomous driving comprehensively utilizes multiple sensors distributed in various parts of the vehicle body to perceive, plan, decide and control the behavior of the vehicle. Autonomous driving represented by Google Waymo uses multiple radars for near, medium and long distance detection, uses multiple cameras to perceive the surrounding environment, uses deep learning methods to identify traffic lights and moving obstacles such as vehicles or pedestrians in front, and focuses on intelligent perception and ultimately controls the behavior of the vehicle through planning and decision-making. In addition to solving the problems of vehicle perception intelligence, information fusion, planning, decision-making, control and other issues, autonomous driving also needs to interact with the environment and surrounding vehicles.

[0003] At present, there is a lack of interaction cognition in unmanned driving. Unmanned vehicles need to autonomously respond to the various uncertainties that often occur during driving, which not only requires vehicle dynamics and various sensors, but also simulates the autonomous prediction and control of the driver's cognitive ability and interacts with the external environment. Therefore, unmanned vehicles need to have sufficient interaction cognitive ability to be accepted by the public.

[0004] Traditional ability evaluation of unmanned driving is mostly focused on the accuracy of perception, stability of control, safety and driving range, etc. as hard indicators, while interaction cognitive ability is more inclined to soft intelligence, which describes the ability of vehicles to understand, predict, communicate and make decisions like humans in complex and dynamic social traffic environments. Traditional performance representation methods of unmanned driving cannot effectively measure the cognitive and interactive ability of vehicles in complex, dynamic and human-machine integrated environments. Therefore, how to effectively represent the interaction cognitive ability of unmanned vehicles has become an important research problem, and there is currently very little research on this.

[0005] A method for evaluating the interactive cognitive ability of an unmanned vehicle is disclosed in Chinese patent application CN118966883A, which includes determining a set of dimensions of the interactive cognitive ability of the unmanned vehicle, and further includes the following steps: determining a set of evaluation objects of the interactive cognitive ability of the unmanned vehicle; and calculating a comprehensive evaluation score C of the interactive cognitive ability of the unmanned vehicle. The method has the disadvantage that the indicators involved in the evaluation still focus on hard indicators in the aspects of perception and control, and the process of interactive cognition is not clearly described, so the interactive cognition is not successfully modeled, and the evaluation effect of the interactive cognitive ability of the unmanned vehicle is weak. SUMMARY

[0006] To solve the above technical problems, the present application provides a method for evaluating the interactive cognitive ability of an unmanned vehicle, which can make the cognitive and interactive processes of the unmanned vehicle system efficient, predictable, and aligned with human intentions. The method quantifies the uncertainty in the decision-making and interactive processes of the unmanned vehicle system using entropy in information theory, and designs and injects stress test scenarios to evaluate the changes in this entropy value, thereby obtaining a comprehensive ability score.

[0007] The present application provides a method for evaluating the interactive cognitive ability of an unmanned vehicle, which includes constructing a hierarchical scenario library and further includes the following steps:

[0008] Step 1: Design stress factors and inject them into each scenario for scenario testing;

[0009] Step 2: Collect test data and extract features during the scenario testing process;

[0010] Step 3: Calculate cognitive entropy and interaction entropy;

[0011] Step 4: Perform comprehensive scoring and level certification.

[0012] Preferably, the scenarios in the hierarchical scenario library are divided into three levels: basic interaction scenarios s1-s4, complex system scenarios s5-s8, and extreme stress scenarios s9-s12.

[0013] In any of the above solutions, preferably, the basic interaction scenarios include:

[0014] S1: unprotected left turn, opposite straight flow;

[0015] S2: merging in dense traffic;

[0016] S3: passing through a pedestrian-vehicle mixed area;

[0017] S4: game-like passing with pedestrians.

[0018] In any of the above solutions, preferably, the complex system scenarios include:

[0019] S5: Ring island passing, need to interact with multiple vehicles game;

[0020] S6: Emergency vehicle avoidance and cooperation;

[0021] S7: Passing sequence negotiation at four-way intersection without signal lights;

[0022] S8: Borrowing passage in construction area, need to coordinate with oncoming vehicles.

[0023] In any of the above solutions, it is preferred that the extreme pressure scenario includes:

[0024] S9: Urban road driving in bad weather;

[0025] S 10 : Encountering aggressive or irrational behavior of other traffic participants;

[0026] S 11 : Partial sensor failure or performance degradation;

[0027] S 12 : Cooperative operation under V2X communication delay or interruption.

[0028] In any of the above solutions, it is preferred that the step 1 includes generating a set of adjustable pressure factors P i for each scenario S j , where 1≤i≤12, 1≤j≤3.

[0029] In any of the above solutions, it is preferred that the pressure factor includes environmental pressure factor P1, traffic participant pressure factor P2 and system internal pressure factor P3.

[0030] In any of the above solutions, it is preferred that the environmental pressure factor P1 includes weather conditions, lighting conditions and road conditions.

[0031] In any of the above solutions, it is preferred that the traffic participant pressure factor P2 includes traffic participant density, uncertainty of other participant behavior and degree of hostility of other participants.

[0032] In any of the above solutions, it is preferred that the system internal pressure factor P3 includes: delay of system perception or decision link, noise of injected sensor data and partial or complete failure of simulation of certain sensors.

[0033] In any of the above solutions, it is preferred that the step 1 further includes applying pressure P i in scenario S x , conducting scenario test, represented as Test(S i , P x ), where Px For a pressure matrix composed of pressure factors of different pressure levels, x is a pressure level, x∈{low, mid, high}, wherein low is a low pressure level, mid is a medium pressure level, and high is a high pressure level.

[0034] In any of the above solutions, preferably, the test data includes ego state data, environment state data, and key event timestamps.

[0035] In any of the above solutions, preferably, the step 3 further includes calculating the cognitive entropy H i and the pressure P x , and the calculation formula is C

[0036]

[0037] wherein N is the number of time points sampled under the scenario S i and the pressure P x , t k is the kth time sampling point, and I(t k ) is the instability of the planned trajectory.

[0038] In any of the above solutions, preferably, the similarity of the planned trajectory at different time points is calculated using the method of trajectory point sequence variance, including:

[0039] Step 301: Plan a trajectory for the future T seconds at time t k , denoted as trajectory , wherein is the position at time t k predicted at time t k +τ;

[0040] Step 302: After a short time interval Δt, generate a new planned trajectory at time t k +Δt;

[0041] Step 303: Align the predictions of trajectory and at the same future time point, and compare and ;

[0042] Step 304: Calculate the instability I(t k ) of all planned trajectories within the time window [t k , t k +Δt], and the formula is

[0043] ​

[0044] Where ||·|| represents the Euclidean distance.

[0045] In any of the above schemes, the method for calculating the interaction entropy preferably includes:

[0046] Step 311: In the exact same scene S i and pressure P x Below, record the behavioral entropy h that affects all participants within the domain at this moment. m (t), as the baseline entropy H baseline ;

[0047] Step 312: Run the unmanned vehicle under test and record the behavioral entropy H of all participants within the influence domain. measured ;

[0048] Step 313: Calculate the interaction entropy H I .

[0049] Preferably, in any of the above schemes, the behavioral entropy h m The formula for calculating (t) is:

[0050]

[0051] Among them, a m (t) represents the acceleration of participant m at time t. Based on its acceleration sequence over a past period Predict the current acceleration a m The probability of (t).

[0052] Preferably, in any of the above schemes, the interaction entropy H I The calculation formula is:

[0053]

[0054] Where M is the average number of participants within the influence domain, and T is the total duration of the scene. This represents the behavioral entropy of participant m at time t in the test scenario. This represents the behavioral entropy of participant m at time t in the baseline scenario.

[0055] In any of the above solutions, step 4 preferably includes the following sub-steps:

[0056] Step 41: Calculate the cognitive entropy H C and interaction entropy H I Perform normalization;

[0057] Step 42: Calculate the Cognitive-Interaction Ability Composite Index (CII(S)) i ,P x );

[0058] Step 43: generating a capability matrix;

[0059] Step 44: performing level certification based on the capability matrix.

[0060] In any of the above solutions, preferably, the cognitive entropy H C The calculation formula for normalization is

[0061]

[0062] wherein, is the maximum value of the cognitive entropy, is the minimum value of the cognitive entropy.

[0063] In any of the above solutions, preferably, the interaction entropy H I The calculation formula for normalization is

[0064]

[0065] wherein, is the maximum value of the interaction entropy, is the minimum value of the interaction entropy.

[0066] In any of the above solutions, preferably, the calculation formula of the cognitive-interaction capability comprehensive index CII(S i , P x ) is

[0067]

[0068] wherein, w1 and w2 are weights, and w1+w2=1.

[0069] In any of the above solutions, preferably, the step 43 comprises traversing all scenarios S i and different levels of stress P x to obtain a multi-dimensional capability matrix,

[0070]

[0071] wherein, P low is a stress matrix composed of low stress level factors, P mid is a stress matrix composed of medium stress level factors, and P high is a stress matrix composed of high stress level factors.

[0072] The present application proposes an interaction cognitive capability evaluation method for unmanned vehicles, which can provide deep insights for R&D personnel, and can also provide a more comprehensive and more reliable unmanned vehicle capability evaluation standard for regulatory agencies, insurance companies and the public. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 This is a flowchart of a preferred embodiment of the interactive cognitive ability evaluation method for autonomous driving according to the present invention. Detailed Implementation

[0074] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0075] Example 1

[0076] like Figure 1 As shown, a method for evaluating interactive cognitive ability for autonomous driving is described. Step 1000 involves constructing a hierarchical scenario library, where the scenarios are divided into three levels: basic interactive scenarios s1-s4, complex system scenarios s5-s8, and extreme stress scenarios s9-s12.

[0077] The basic interaction scenarios include:

[0078] S1: Unprotected left turn, with oncoming traffic flowing straight;

[0079] S2: Lane merging in heavy traffic;

[0080] S3: Passage through areas where pedestrians and vehicles share the same road;

[0081] S4: Game-like passage with pedestrians.

[0082] The complex system scenarios include:

[0083] S5: Roundabout passage requires interaction and strategic maneuvering with multiple vehicles;

[0084] S6: Emergency vehicle avoidance and cooperation;

[0085] S7: Negotiation of traffic order at four-way intersections without traffic lights;

[0086] S8: When using the road in the construction area, coordination with oncoming vehicles is required.

[0087] The extreme stress scenarios include:

[0088] S9: Driving on city roads in inclement weather;

[0089] S 10 Encountering aggressive or irrational behavior from other road users;

[0090] S 11 Some sensors have failed or their performance has deteriorated;

[0091] S 12 Cooperative operation under V2X communication delays or interruptions.

[0092] performing step 1100, designing pressure factors, injecting the pressure factors into each scenario, and conducting scenario testing, including for each scenario S i generating a set of adjustable pressure factors P j where 1≤i≤12, 1≤j≤3; under scenario S i applying pressure P x conducting testing, denoted as Test(S i , P x ), where P x is a pressure matrix composed of pressure factors of different pressure levels, x is a pressure level, x∈{low, mid, high}, where low is a low pressure level, mid is a mid pressure level, and high is a high pressure level.

[0093] The pressure factors include environmental pressure factors, traffic participant pressure factors, and system internal pressure factors.

[0094] The environmental pressure factors include weather conditions, lighting conditions, and road surface conditions.

[0095] The traffic participant pressure factors include traffic participant density, uncertainty of other participant behavior, and hostility level of other participants.

[0096] The system internal pressure factors include delay of system perception or decision link, noise level of sensor data, and partial or complete failure of simulating specific sensors.

[0097] performing step 1200, collecting testing data and extracting features during testing, the testing data including ego vehicle state data, environmental state data, and key event timestamps.

[0098] performing step 1300, calculating cognitive entropy and interaction entropy, including calculating cognitive entropy H i under scenario S and pressure C , the calculation formula being

[0099]

[0100] where N is the number of time points sampled under scenario S i and pressure P x , t k is the kth time sampling point, and I(t k ) is the instability of the planned trajectory.

[0101] calculating the similarity of the planned trajectory at different time points using the method of trajectory point sequence variance, the higher the similarity, the lower the instability of the trajectory, including:

[0102] Execute step 1301, in t k The trajectory is planned for the next T seconds, and represented as a trajectory. ,in, For in t k The predicted time t k The position at time +τ;

[0103] Execute step 1302, after a short time interval Δt, at t k A new planned trajectory is generated at time +Δt. ;

[0104] Execute step 1303 to align the trajectory. and Comparison of predictions at the same future point in time and ;

[0105] Execute step 1304 to calculate within the time window [t] k ,t k Within +Δt], the instability I(t) of all planned trajectories k The formula is:

[0106]

[0107] Where ||·|| represents the Euclidean distance.

[0108] The method for calculating the interaction entropy includes:

[0109] Execute step 1311 in the exact same scenario S i and pressure P x Below, record the behavioral entropy h that affects all participants within the domain at this moment. m (t), as the baseline entropy H baseline The behavioral entropy h m The formula for calculating (t) is:

[0110]

[0111] Among them, a m (t) represents the acceleration of participant m at time t. Based on its acceleration sequence over a past period Predict the current acceleration a m The probability of (t).

[0112] Execute step 1312, run the unmanned vehicle under test, and record the behavioral entropy H of all participants within the influence domain. measured .

[0113] Perform step 1313 to calculate the interaction entropy H. I, the interaction entropy H I The calculation formula is

[0114]

[0115] Wherein, M is the average number of participants in the influence domain, T is the total duration of the scene, Indicates the behavior entropy of participant m at time t in the test scene, Indicates the behavior entropy of participant m at time t in the baseline scene.

[0116] Step 1400 is performed, and comprehensive scoring and level certification are performed, and the step 4 includes the following sub-steps:

[0117] Step 1410 is performed, and the cognitive entropy H C And the interaction entropy H I Normalization is performed on the cognitive entropy H C The calculation formula for normalizing the cognitive entropy H

[0118]

[0119] Wherein, The maximum value of cognitive entropy, The minimum value of cognitive entropy.

[0120] The calculation formula for normalizing the interaction entropy H I

[0121]

[0122] Wherein, The maximum value of interaction entropy, The minimum value of interaction entropy.

[0123] Step 1420 is performed, and the cognitive-interactive ability comprehensive index CII(S i , P x ) is calculated, and the calculation formula of the cognitive-interactive ability comprehensive index CII(S i , P x ) is

[0124]

[0125] Wherein, w1 and w2 are weights, and w1+w2=1.

[0126] Step 1430 is performed, and the ability matrix is generated, including traversing all scenes S i And different levels of stress P x A multi-dimensional ability matrix is obtained,

[0127] ​

[0128] where P low is a pressure matrix composed of low pressure level factors, P mid is a pressure matrix composed of medium pressure level factors, P high is a pressure matrix composed of high pressure level factors.

[0129] Performing step 1440, based on the capability matrix, performing level authentication.

[0130] Embodiment two

[0131] The present application proposes an interactive cognitive ability evaluation method for unmanned driving, so that the cognitive and interactive process of the unmanned driving system is efficient, predictable, and aligned with human intention. The present application quantifies the uncertainty in the decision-making and interactive process through entropy in information theory, and designs and injects pressure test scenarios to evaluate the change of entropy value, thereby obtaining the comprehensive ability score, so that the interactive cognitive process of the unmanned driving system is efficient and predictable. The key steps are:

[0132] 1. Scene: No longer rely on fixed routes or random events, but build a series of complex scene library with clear test objectives and progressive levels. These scenarios aim to stimulate the extreme ability of the unmanned driving system in perception, prediction, decision-making, planning, and interaction.

[0133] 2. Pressure injection: On the basis of the scene, dynamically inject pressure factors, including abnormal behavior of other traffic participants, adverse weather, communication delay, sensor noise, etc., to simulate challenging situations in the real world.

[0134] 3. Quantification of entropy: The behavior of the unmanned driving system is divided into cognitive entropy and interactive entropy. Cognitive entropy measures the uncertainty of the internal decision-making process of the system. A system with strong cognitive ability should be able to quickly converge to an optimal or suboptimal decision in the face of complex situations, and its cognitive entropy should be lower. Interactive entropy measures the uncertainty of the system's interaction with the external environment. A system with strong interaction ability should be able to clearly understand the behavior of other traffic participants, thereby guiding predictable reactions and reducing the entropy of the entire interactive system.

[0135] Specifically includes:

[0136] 1. Build a hierarchical scene library

[0137] Establish a standardized and reproducible test scene library. The scenes are divided into three levels:

[0138] L1 - Basic interaction scene: Test basic interaction rules.

[0139] S1: Left turn without protection, opposite straight traffic flow;

[0140] S2: Lane merging in dense traffic flow;

[0141] S3: Passage in mixed pedestrian and vehicle area (e.g. neighborhood gate).

[0142] S4: "Game" passage with pedestrians (e.g. before a pedestrian crossing, pedestrians hesitate).

[0143] L2 - Complex coordination scenarios: Test multi-agent coordination and intent understanding.

[0144] S5: Roundabout passage, need to interact with multiple vehicles.

[0145] S6: Emergency vehicle avoidance and cooperation.

[0146] S7: Passage order negotiation at four-way signal-free intersection.

[0147] S8: Borrowing passage in construction area, need to coordinate with oncoming vehicles.

[0148] L3 - Extreme stress scenarios: Test system robustness in extreme and abnormal situations.

[0149] S9: Urban road driving in bad weather (heavy rain / snow).

[0150] S 10 : Encounters with aggressive and irrational behavior of other traffic participants, including malicious lane changing, sudden pedestrian crossing, etc.

[0151] S 11 : Partial sensor failure or performance degradation.

[0152] S 12 : Cooperative operation under V2X communication delay or interruption.

[0153] 2. Design and inject stress factors

[0154] For each scenario S i , 1≤i≤12, define a set of adjustable stress factors P j , 1≤j≤3.

[0155] Environmental stress factors

[0156] Weather conditions (sunny / rain / snow / fog), value [0,1], 0 is best, 1 is worst.

[0157] Lighting conditions (day / dusk / night / backlight), value [0,1].

[0158] Road conditions (dry / slippery / icy), value [0,1].

[0159] Traffic participant stress factors

[0160] Density of traffic participants (vehicles / m or people / m2).

[0161] Uncertainty of other participants' behavior, value [0, 1].

[0162] Hostility level of other participants, value [0, 1].

[0163] Internal stress factor of the system

[0164] Simulated delay of system perception or decision link.

[0165] Noise level of sensor data.

[0166] Simulate partial or complete failure of specific sensors.

[0167] Divide the stress level into high, medium, and low, represented as low, mid, and high, respectively. The level is represented by variable x. The stress matrix composed of stress factors at different stress levels is P x .

[0168] A specific test can be represented as: Test(S i , P x ), that is, test under scenario S i , with stress P x .

[0169] 3. Data collection and feature extraction

[0170] During the test, the following data is recorded.

[0171] Self-vehicle state data

[0172] Position, speed, acceleration, heading angle, angular velocity.

[0173] Planned path (sequence of trajectory points in the future 5-10 seconds).

[0174] Decision state (following, changing lanes, braking, waiting).

[0175] Interaction signals (turn signal, HMI screen display information, voice prompt).

[0176] Environmental state data

[0177] Position, speed, acceleration, type (car / person / bicycle) of all other traffic participants.

[0178] Traffic signal, sign status.

[0179] Key event timestamp

[0180] Time of conflict occurrence.

[0181] Interaction success moment.

[0182] Moment when system decision changes significantly.

[0183] 4. Calculate cognitive-interaction entropy

[0184] 4.1 Calculation of cognitive entropy

[0185] Cognitive entropy measures the degree of hesitation of the ego vehicle at decision points, quantified by analyzing the evolution of its planned trajectory.

[0186] Calculate cognitive entropy under scenario S i and stress P x , 1≤i≤12. t k is the kth time sampling point. Within the time window [t k , t k +Δt], the dynamic change degree of the planned trajectory of the system for the future is calculated. The more drastic and frequent the changes, the more unstable the internal cognitive model, and the higher the cognitive entropy. The method of variance of trajectory point sequence is used to measure the similarity of planned trajectories at different time points, and the specific calculation process is as follows.

[0187] 1. At time t k , the system plans a trajectory for the future T seconds, represented as a point sequence , where is the position predicted at t k +τ at t k +τ.

[0188] 2. After a short time interval Δt, at t k +Δt, the system generates a new planned trajectory .

[0189] 3. Compare and , align the predictions of the two trajectories at the same future time point.

[0190] 4. Calculate the instability I(t k ) of all planned trajectories within the time window [t k , t k +Δt].

[0191] The formula for calculating trajectory instability is as follows:

[0192]

[0193] where ||·|| is the Euclidean distance.

[0194] In the entire test scenario, the cognitive entropy H Cis the expected and normalized trajectory instability. The formula of cognitive entropy is as follows:

[0195]

[0196] H C The lower, the more decisive and stable the system decision is.

[0197] 4.2 Calculation of Interaction Entropy

[0198] Interaction entropy measures the influence of the ego vehicle behavior on the surrounding traffic environment. If the ego vehicle behaves clearly and predictably, other participants can more easily predict its intention and thus make stable responses, and the interaction entropy of the whole system is low. The influence domain represents a circular area centered on the ego vehicle with a radius of R, and other traffic participants within the area are considered to be directly affected. For each other participant m within the influence domain, the randomness of its behavior is measured by the change in its acceleration, called the behavior entropy of the other participant. A vehicle that frequently and sharply accelerates and decelerates has a higher behavior entropy. Interaction entropy is calculated by observing the change in the randomness of the trajectories of other participants within the influence domain before and after the ego vehicle behavior.

[0199] The behavior entropy h m (t) of participant m at time t can be defined as:

[0200]

[0201] Here a m (t) is the acceleration of participant m at time t, is the probability of predicting the current acceleration a (t) from the acceleration sequence of participant m in the past time period. The more accurate the prediction (the higher the probability), the lower the entropy h m (t) of participant m, indicating that the behavior is more predictable.

[0202] Interaction entropy represents the amount of increase in the behavior entropy of other participants due to the ego vehicle behavior. To isolate the influence of the ego vehicle, a baseline is first introduced. In the same scenario S i and stress P x , the baseline scenario replaces the ego vehicle under test with a perfect and behavior-predictable hypothetical vehicle, and records the behavior entropy of all participants within the influence domain as the baseline entropy H baseline . The ego vehicle under test is run in the real scenario, and the behavior entropy H measured of all participants within the influence domain is recorded. The interaction entropy H I (S i , P x ) is calculated as follows:

[0203]

[0204] where M is the average number of participants in the domain, T is the total duration of the scenario, represents the behavior entropy of participant m at time t in the test scenario, represents the behavior entropy of participant m at time t in the baseline scenario. I (S i , P x ) is larger, the more the behavior of the vehicle under test increases the confusion of other participants, and the poorer the interaction ability.

[0205] H I (S i , P x ) ≈ 0: the interaction effect of the vehicle under test is equivalent to that of a perfectly rational vehicle, and the interaction ability is excellent.

[0206] H I (S i , P x ) > 0: the behavior of the vehicle under test increases the confusion of other participants, and the interaction ability is poor. The larger the value, the more confusing the behavior, which may cause other vehicles to suddenly stop or pedestrians to hesitate.

[0207] H I (S i , P x ) < 0 (ideal case): the behavior of the vehicle under test is more conducive to guiding traffic than that of a perfectly rational vehicle, making the system smoother. This usually occurs when the vehicle can actively and clearly communicate its intentions, such as turning on the turn signal at the right time or explicitly informing pedestrians through the HMI screen that "you go first".

[0208] 5. Comprehensive score

[0209] Combine cognitive entropy and interaction entropy to get a comprehensive cognitive- interaction ability index.

[0210] 5.1 Entropy value normalization

[0211] Since the entropy values under different scenarios and pressures are of different dimensions, H C and H I need to be normalized to map to the [0, 1] interval.

[0212]

[0213]

[0214] is the maximum value of cognitive entropy, is the minimum value of cognitive entropy. is the maximum value of interaction entropy, is the minimum value of interaction entropy.

[0215] 5.2 Calculate the comprehensive index

[0216] The formula for calculating the cognitive-interaction ability index is as follows:

[0217]

[0218] where w1 and w2 are weights, w1+w2=1. Adjust the weights according to the scene type, w2 can be higher in interaction-intensive scenes. The value range of CII is [0,1], the closer to 1 represents the stronger ability.

[0219] 5.3 Generate the ability matrix.

[0220] For an unmanned system, traverse all scenes and different levels of stress , get a multi-dimensional ability portrait matrix.

[0221]

[0222] This matrix can intuitively show which aspects of the system are strong and which aspects are weak.

[0223] The present application provides a more comprehensive and more reliable evaluation method for the cognitive interaction ability of unmanned systems for researchers, regulatory agencies, and unmanned companies.

[0224] 1. It goes beyond the binary evaluation of whether an accident occurs, and goes deep into the quality level of decision-making and interaction, measures the intelligence of the system, and realizes the transition from passive safety to active cognition.

[0225] 2. Through the mathematical tool of entropy, the concepts of fuzzy interaction friendliness and decision-making decisiveness become quantifiable and comparable, facilitating horizontal comparison between different systems.

[0226] 3. Scene and stress-oriented, the evaluation is more targeted and can accurately locate the system's ability short board, providing a clear direction for algorithm optimization.

[0227] 4. The calculation formula of scene library, stress factor and entropy provides a unified and scientific evaluation framework for the industry, with standardization potential.

[0228] Example three

[0229] Based on the ability matrix, set the certification standards for level recognition. For example:

[0230] L4-urban commuting level: requires CII>0.85 in all L1, L2 scenes and low, medium stress; CII>0.7 in L3 low stress.

[0231] L4 - full duty class: requires CII > 0.8 in all scenarios and all pressure levels. In order to better understand the present application, the above is described in detail in combination with the specific embodiments of the present application, but is not a limitation on the present application. Any simple modification made to the above embodiments in accordance with the technical essence of the present application still belongs to the scope of the technical solutions of the present application. In the specification, each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For system embodiments, since they basically correspond to method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

Claims

1. A method for evaluating interactive cognitive abilities for autonomous driving, comprising constructing a hierarchical scene library, characterized in that, It also includes the following steps: Step 1: Design stress factors, inject the stress factors into each scenario, and conduct scenario testing; Step 2: During the scenario testing process, collect test data and extract features; Step 3: Calculate cognitive entropy - interaction entropy; Step 4: Conduct comprehensive scoring and level certification.

2. The interactive cognitive ability evaluation method for autonomous driving as described in claim 1, characterized in that, Step 1 includes providing each scenario S i Generate a set of adjustable pressure factors P j , where 1≤i≤12, 1≤j≤3.

3. The interactive cognitive ability evaluation method for autonomous driving as described in claim 2, characterized in that, Step 1 also includes in scenario S i Apply pressure P x To perform a test, denoted as Test(S) i , P x ), where P x This is a pressure matrix composed of pressure factors of different pressure levels, where x represents the pressure level.

4. The interactive cognitive ability evaluation method for autonomous driving as described in claim 3, characterized in that, Step 3 also includes in scenario S i and pressure P x Calculate cognitive entropy H C The calculation formula is: ； Where N is the value in scene S i and pressure P x Number of downsampling time points, t k For the k-th time sampling point, I(t) k ) represents the instability of the planned trajectory.

5. The interactive cognitive ability evaluation method for autonomous driving as described in claim 4, characterized in that, The similarity of planned trajectories at different time points is calculated using the variance of trajectory point sequences, including: Step 301: In t k The trajectory is planned for the next T seconds, and represented as a trajectory. ,in, For in t k The predicted time t k The position at time +τ; Step 302: After a short time interval Δt, at t k A new planned trajectory is generated at time +Δt. ; Step 303: Align the trajectory and Comparison of predictions at the same future point in time and ; Step 304: Calculate within the time window [t] k ,t k Within +Δt], the instability I(t) of all planned trajectories k The formula is: ； Where ||·|| represents the Euclidean distance.

6. The interactive cognitive ability evaluation method for autonomous driving as described in claim 5, characterized in that, The method for calculating the interaction entropy includes: Step 311: In the exact same scene S i and pressure P x Below, record the behavioral entropy h that affects all participants within the domain at this moment. m (t), as the baseline entropy H baseline ; Step 312: Run the unmanned vehicle under test and record the behavioral entropy H of all participants within the influence domain. measured ; Step 313: Calculate the interaction entropy H I .

7. The interactive cognitive ability evaluation method for autonomous driving as described in claim 6, characterized in that, The behavioral entropy h m The formula for calculating (t) is: ； Among them, a m (t) represents the acceleration of participant m at time t. Based on its acceleration sequence over a past period Predict the current acceleration a m The probability of (t).

8. The interactive cognitive ability evaluation method for autonomous driving as described in claim 7, characterized in that, The interaction entropy H I The calculation formula is ； Where M is the average number of participants within the influence domain, and T is the total duration of the scene. This represents the behavioral entropy of participant m at time t in the test scenario. This represents the behavioral entropy of participant m at time t in the baseline scenario.

9. The interactive cognitive ability evaluation method for autonomous driving as described in claim 8, characterized in that, Step 4 includes the following sub-steps: Step 41: Calculate the cognitive entropy H C and interaction entropy H I Perform normalization; Step 42: Calculate the Cognitive-Interaction Ability Composite Index (CII(S)) i ,P x ); Step 43: Generate the capability matrix; Step 44: Perform level certification based on the capability matrix.

10. The interactive cognitive ability evaluation method for autonomous driving as described in claim 9, characterized in that, The cognitive entropy H C The formula for normalization is: ； in, This represents the maximum value of cognitive entropy. This represents the minimum value of cognitive entropy.

Images