Intelligent driving high-discrimination evaluation automatic generation method based on historical test data

Abstract

The application relates to the technical field of intelligent driving and discloses an intelligent driving high-discrimination evaluation automatic generation method based on historical test data, which comprises the following steps: acquiring historical kinematics test data and determining a candidate kinematics segment; determining disturbance parameters and kinematics evaluation indexes based on the candidate kinematics segment, constructing a candidate evaluation method, and generating a disturbance test segment according to a plurality of disturbance parameter values; causing the same group of disturbance test segments to act on at least two measured objects, and acquiring kinematics evaluation index results; determining version discrimination according to the result difference under the same disturbance test segment, and determining index sensitivity according to the change of the index result with the disturbance parameter value; and determining a high-discrimination evaluation method based on the version discrimination and the index sensitivity. The scheme solves the problem that the version discrimination capability and the evaluation index response capability are difficult to be considered in intelligent driving evaluation.

Description

Technical Field

[0001] This specification relates to the field of intelligent driving technology, and in particular to an automatic generation method for high-discrimination intelligent driving evaluation based on historical test data. Background Technology

[0002] In the research and development and verification of autonomous driving systems, it is usually necessary to compare and evaluate different algorithm versions, control strategies, or parameter configurations. Road testing, closed-course testing, and automated testing accumulate a large amount of historical kinematic test data. Therefore, how to utilize this data to generate evaluation methods that can effectively distinguish different test subjects has become an important technical problem in intelligent driving evaluation.

[0003] Existing methods typically compare the results of a single test of different autonomous driving solutions under fixed test conditions, rarely adjusting conditions such as initial distance, relative speed, or duration of risk for the same driving segment at multiple levels, or having each autonomous driving solution repeatedly tested under the same multiple sets of conditions. Therefore, it is difficult to determine whether the differences in results between different autonomous driving solutions are significant, or whether the evaluation indicators can effectively respond to changes in test conditions, thus making it difficult to select an evaluation method that can reliably distinguish between different autonomous driving solutions.

[0004] While some testing methods can differentiate between different algorithm versions under specific conditions, this difference may not be maintained after the testing conditions change. Similarly, while some evaluation metrics can change stably with testing conditions, they may not effectively distinguish between different algorithm versions. Therefore, existing technologies struggle to balance version differentiation capabilities with test result stability, necessitating an automated method for generating highly discriminative evaluations of intelligent driving based on historical test data to address these issues. Summary of the Invention

[0005] In view of this, the present invention aims to propose an automatic generation method for high-discrimination evaluation of intelligent driving based on historical test data, in order to solve the problem that existing intelligent driving evaluation methods are difficult to simultaneously consider the differences in results between different test subjects and the stability of test results as test conditions change, thus making it difficult to automatically select high-discrimination evaluation methods.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for automatically generating high-discrimination evaluations of intelligent driving based on historical test data includes: S1: Obtain historical kinematic test data and determine candidate kinematic segments based on the historical kinematic test data; S2: Determine perturbation parameters and kinematic evaluation indicators based on the candidate kinematic segments, construct candidate evaluation methods based on the candidate kinematic segments, the perturbation parameters, and the kinematic evaluation indicators, and generate multiple perturbation test segments according to multiple perturbation parameter values ​​of the perturbation parameters; S3: Apply the same set of disturbance test segments to at least two test objects respectively, and obtain the kinematic evaluation index results of each test object under each disturbance test segment; S4: Based on the differences between the kinematic evaluation index results of different test objects under the same perturbation test segment, determine the version discrimination of the candidate evaluation method, and based on the degree of change of the kinematic evaluation index results with the perturbation parameter value, determine the index sensitivity of the kinematic evaluation index. S5: Based on the version discrimination and the index sensitivity of the kinematic evaluation index, determine the high discrimination evaluation method from the candidate evaluation methods.

[0007] Furthermore, S1 includes: S1-1: Obtain continuous historical kinematic test data, and construct derived kinematic features based on the historical kinematic test data; S1-2: Determine the risk evolution stage based on the changes in the derived kinematic characteristics over a continuous time range; S1-3: Based on the start and end times of the risk evolution stage, extract the candidate kinematic segments from the historical kinematic test data.

[0008] Furthermore, S2 includes: S2-1: Arrange the historical kinematic test data and derived kinematic features in the candidate kinematic segments in chronological order to obtain candidate kinematic time series data; S2-2: Input the candidate kinematic time series data into the large model, and use the large model to perform semantic analysis on the risk evolution process and vehicle control process in the candidate kinematic segments to obtain a behavioral risk evolution description including the risk evolution stage and the key kinematic change features corresponding to each risk evolution stage; S2-3: Determine perturbation parameters based on the behavioral risk evolution description, determine kinematic evaluation indicators based on the candidate kinematic segments and their derived kinematic features, and construct the candidate evaluation method based on the candidate kinematic segments, the perturbation parameters, and the kinematic evaluation indicators; adjust the candidate kinematic segments according to multiple perturbation parameter values ​​of the perturbation parameters to generate the multiple perturbation test segments.

[0009] Furthermore, S2-3 includes: S2-3-1: Determine the risk evolution stage to be disturbed based on the behavioral risk evolution description, and determine the disturbance parameters corresponding to the risk evolution stage to be disturbed based on the preset mapping relationship between the risk evolution stage and the disturbance parameters; S2-3-2: Determine the perturbation direction of the perturbation parameter based on the key kinematic change characteristics corresponding to the risk evolution stage to be perturbed in the behavioral risk evolution description; S2-3-3: Under the constraints of vehicle kinematic realizability and kinematic temporal continuity, the disturbance range of the disturbance parameter is determined according to the disturbance parameter and the disturbance direction, and a constrained counterfactual disturbance template is generated based on the disturbance parameter, the disturbance direction and the disturbance range; S2-3-4: Adjust the candidate kinematic segments according to the multiple perturbation parameter values ​​in the constrained counterfactual perturbation template to generate the multiple perturbation test segments.

[0010] Furthermore, in S4, for any two tested objects, the first... The calculation logic for the single-index difference corresponding to each kinematic evaluation index is as follows:

[0011] In the formula, Indicates the first The degree of difference between individual indicators corresponding to each kinematic evaluation indicator This indicates the number of perturbation test segments. Indicates the first tested object in the th... The first perturbation test segment under the first Results of kinematic evaluation indicators Indicates that the second tested object is in the first... The first perturbation test segment under the first Results of kinematic evaluation indicators Indicates the first The normalized range corresponding to each kinematic evaluation index This indicates a preset positive number.

[0012] Furthermore, in S4, the version discrimination is determined based on the single-index difference corresponding to multiple kinematic evaluation indicators, and its calculation logic is as follows:

[0013]

[0014] In the formula, This indicates the version differentiation. This indicates the number of kinematic evaluation metrics used in the version discrimination calculation. Indicates the first The discrimination weights corresponding to each kinematic evaluation index.

[0015] Further, in S4, determining the sensitivity of any kinematic evaluation index for any tested object includes: arranging multiple perturbation parameter values ​​in numerical order according to their values, and determining the sensitivity of the kinematic evaluation index based on the response degree of the kinematic evaluation index results corresponding to adjacent perturbation parameter values. The calculation logic for the sensitivity of the kinematic evaluation index is as follows:

[0016] In the formula, This indicates the sensitivity of the kinematic evaluation index. This indicates the number of perturbation test segments that correspond one-to-one with each of the perturbation parameter values. and They represent the adjacent numbers arranged in the numerical order. The first disturbance parameter value and the second One disturbance parameter value, and These respectively represent the tested object at the 1st epoch. The first disturbance parameter value and the second The kinematic evaluation index results obtained under the perturbation test segment corresponding to each perturbation parameter value This indicates a preset positive number.

[0017] Furthermore, S5 includes: S5-1: Compare the version discrimination with the preset version discrimination threshold, and compare the index sensitivity of the kinematic evaluation index corresponding to each tested object with the preset index sensitivity threshold respectively; S5-2: When the version differentiation reaches the preset version differentiation threshold, the index sensitivity of the kinematic evaluation index corresponding to each tested object reaches the preset index sensitivity threshold, and the corresponding disturbance test segment satisfies the vehicle kinematics realizability constraint and the kinematics temporal continuity constraint, the combination of the corresponding candidate kinematic segment, disturbance parameter and the kinematic evaluation index is determined as the target evaluation combination. S5-3: Generate the high-discrimination assessment method based on the target assessment combination.

[0018] Furthermore, in S5-3, generating the high-discrimination evaluation method includes: S5-3-1: Extract the candidate kinematic segments, the perturbation parameters, and the kinematic evaluation indicators from the target evaluation combination; determine the range of perturbation parameters based on the multiple perturbation parameter values ​​corresponding to the perturbation parameters; and determine the test execution conditions and judgment conditions. S5-3-2: Associate the candidate kinematic segments, the perturbation parameter range, the test execution conditions, the kinematic evaluation index, and the judgment conditions to generate evaluation method configuration data that can be called by at least one of the playback test system and the simulation test system.

[0019] The principles and advantages of this invention are: 1. By processing historical kinematic test data, the risk evolution stages are determined based on the changes in derived kinematic characteristics over time, and candidate kinematic segments are identified according to the start and end positions of each risk evolution stage. The kinematic time-series data corresponding to the candidate kinematic segments are input into the large model to generate a behavioral risk evolution description that includes the risk evolution stages and key kinematic change characteristics. Based on this, the perturbation parameters, perturbation direction, and perturbation parameter range are determined, and multiple perturbation test segments are generated under vehicle kinematic realizability constraints and kinematic time-series continuity constraints. This allows for the automatic generation of executable test segments with different test conditions based on historical test data, reducing the reliance of evaluation methods on manual segment selection and parameter adjustment.

[0020] 2. By applying the same set of perturbation test segments to at least two test subjects, the kinematic evaluation index results of each test subject under the same perturbation test segments are obtained. The single-index difference and version discrimination are determined based on the differences in index results between different test subjects. Simultaneously, the index sensitivity is determined according to the degree of change in the evaluation index results corresponding to adjacent perturbation parameter values, in order of magnitude of the perturbation parameter values. Therefore, it is possible to determine whether the candidate evaluation method can effectively distinguish different test subjects and whether the adopted kinematic evaluation index will produce a significant response to changes in perturbation parameters.

[0021] 3. By comparing version discrimination and index sensitivity with corresponding thresholds, and considering whether the disturbance test segments meet the vehicle kinematics realizability constraints and kinematic temporal continuity constraints, a target evaluation combination consisting of candidate kinematic segments, disturbance parameters, and kinematic evaluation indices is determined. Furthermore, the range of disturbance parameters, test execution conditions, and judgment conditions are determined, and evaluation method configuration data that can be invoked by at least one of the playback test system and simulation test system is generated. Thus, a highly discriminative evaluation method that can be repeatedly executed can be selected and output from multiple candidate evaluation methods. Attached Figure Description

[0022] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1This is a flowchart illustrating the automatic generation method for high-discrimination evaluation of intelligent driving based on historical test data. Detailed Implementation

[0023] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0024] As indicated in this specification and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0025] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0026] The following detailed explanation illustrates the specific implementation methods: like Figure 1 As shown, the automatic generation method for high-discrimination intelligent driving evaluation based on historical test data includes: S1: Obtain historical kinematic test data and determine candidate kinematic segments based on the historical kinematic test data.

[0027] Historical kinematic test data refers to the vehicle's apparent motion state data collected by the vehicle testing system or data recording system during road testing, closed-track testing, or automated testing.

[0028] In this embodiment, historical kinematic test data may include one or more of the following: vehicle speed, vehicle longitudinal acceleration, vehicle lateral acceleration, inter-vehicle distance, and time. The data are correlated according to timestamps to form a continuous data sequence.

[0029] After acquiring historical kinematic test data, the following steps can be performed: time alignment, outlier handling, and unit unification. Time alignment is used to map data from different acquisition frequencies to the same time axis; outlier handling is used to remove abrupt changes that exceed the reasonable range of vehicle motion; and unit unification is used to convert speed, acceleration, distance, and time into preset unified units.

[0030] S1 includes: S1-1: Obtain continuous historical kinematic test data and construct derived kinematic features based on the historical kinematic test data.

[0031] In this embodiment, the derived kinematic features include one or more of the following: longitudinal acceleration rate of change, lateral acceleration rate of change, distance rate of change, relative approach velocity, collision time, and time interval.

[0032] Let the first Each sampling time is The time interval between adjacent sampling times is The vehicle's longitudinal acceleration is The vehicle's lateral acceleration is The distance between vehicles is The vehicle speed is The rate of change of longitudinal acceleration is determined by the following formula:

[0033] In the formula, Indicates the first Each sampling time The corresponding rate of change of longitudinal acceleration; Indicates the first The longitudinal acceleration of the vehicle at each sampling time; Indicates the previous sampling time The vehicle's longitudinal acceleration; This represents the time interval between two adjacent sampling times, i.e. .

[0034] Among them, vehicle longitudinal acceleration The corresponding sampling time can be directly read from the longitudinal acceleration field and time field in the historical kinematic test data.

[0035] This formula calculates the ratio of the difference in longitudinal acceleration between adjacent sampling times to the time interval, thus obtaining the degree of change in longitudinal acceleration over time. The larger the absolute value, the greater the longitudinal acceleration of the vehicle. The more pronounced the changes are within the corresponding time period.

[0036] The rate of change of lateral acceleration is determined by the following formula:

[0037] In the formula, Indicates the first Each sampling time The corresponding rate of change of lateral acceleration; Indicates the first The lateral acceleration of the vehicle at each sampling time; Indicates the previous sampling time The vehicle's lateral acceleration; This represents the time interval between two adjacent sampling times.

[0038] The vehicle's lateral acceleration and its corresponding sampling time can be obtained from the lateral acceleration field and time field in the historical kinematic test data.

[0039] This formula calculates the ratio of the difference in lateral acceleration between adjacent sampling times to the time interval, thus obtaining the degree of change in the vehicle's lateral acceleration over time. The larger the absolute value, the more obvious the change in the vehicle's lateral motion state.

[0040] The rate of change of distance is determined by the following formula:

[0041] In the formula, Indicates the first Each sampling time The corresponding rate of change of distance; Indicates the first The distance between the vehicle and the target object at each sampling time; This indicates the distance between the current vehicle and the target object at the previous sampling time. This represents the time interval between two adjacent sampling times.

[0042] The distance between vehicles can be obtained from the distance between vehicles field in the historical kinematic test data, or it can be calculated based on the position data of the vehicle and the target object at the same sampling time.

[0043] The rate of change of vehicle distance over time is obtained by calculating the ratio of the distance difference between vehicles at adjacent sampling times to the time interval. When the distance is less than 0, it indicates that the distance between the vehicle and the target is decreasing. When the value is greater than 0, it indicates that the distance between the two is increasing.

[0044] The relative approximation velocity is determined by the following formula:

[0045] In the formula, Indicates the first The relative approximation speed of the vehicle relative to the target object at each sampling time; due to the shortening distance... If the value is negative, by taking the opposite of the rate of change of distance, the relative approximation velocity corresponding to the shortening of distance becomes positive. Therefore, when... When the value is greater than 0, it indicates that the distance between the vehicle and the target is decreasing, and The larger the value, the faster the two approximate each other; when... When ≤0, it means that there is no approximation trend between the two.

[0046] The collision time is determined by the following formula:

[0047] In the formula, Indicates the first Collision time at each sampling moment; This indicates the distance between the current vehicle and the target object at the sampling time. This indicates the relative approximation velocity at that sampling moment; This represents a preset minimum positive number used to prevent the denominator from being zero.

[0048] in, It can be obtained from historical kinematic test data or calculated from location data. Calculated based on the rate of change of distance. When the distance between the vehicles is greater than 0, it indicates that the vehicle is approaching the target object. At this point, dividing the distance between the vehicles by the relative approach speed yields the estimated time required for a collision based on the current approach trend. When the value is ≤0, it indicates that there is no tendency for the two to approach each other, and the collision time is set to positive infinity. The smaller the value, the higher the current collision risk.

[0049] The time interval is determined according to the following formula:

[0050] In the formula, For the first The time interval between each sampling moment; This indicates the distance between the current vehicle and the target object at the sampling time. This indicates the vehicle's speed at the time of sampling; This represents a preset minimum positive number used to prevent the denominator from being zero when the vehicle's speed is zero.

[0051] The vehicle speed and inter-vehicle distance can be obtained from the vehicle speed field and inter-vehicle distance field in the historical kinematic test data, respectively.

[0052] This formula calculates the time required for the vehicle to travel to the target's current location at its current speed by dividing the distance between the vehicles by the vehicle's speed. The smaller the value, the lower the safety margin for following the target vehicle.

[0053] S1-2: Determine the risk evolution stage based on the changes in derived kinematic characteristics over a continuous time range.

[0054] The risk evolution stage is used to describe the temporal changes of a candidate kinematic process from risk formation to risk resolution.

[0055] As an example only, the risk evolution phases can include the risk approach phase, the control response phase, the risk mitigation phase, and the recovery phase.

[0056] The risk approach phase can be determined based on one or more of the following: the distance between vehicles continues to shorten, the relative approach speed is positive, the collision time continues to decrease, or the time-distance continues to decrease. The control response phase can be determined based on significant changes in longitudinal acceleration, lateral acceleration, the rate of change of longitudinal acceleration, or the rate of change of lateral acceleration. The risk mitigation phase can be determined based on the cessation of the shortening of the distance between vehicles and the recovery of the collision time or the time-distance. The recovery phase can be determined based on the gradual recovery of vehicle speed, acceleration, or distance between vehicles to stability after the risk is eliminated.

[0057] In this embodiment, the threshold for each risk evolution stage can be preset based on historically labeled samples, test task requirements, or the permissible range of vehicle kinematics. The threshold is only used to determine the stage boundary and is not limited to a fixed value.

[0058] Furthermore, the risk exposure duration is used to characterize the duration during which a vehicle is in a state with low safety margin or high approach risk, and it can be determined according to the following formula:

[0059] In the formula, This indicates the duration of risk exposure corresponding to the candidate kinematic segment; Indicates the number of sampling times included in the statistics; Indicates the first The determination value for whether a sampling moment is in a state of low safety margin or high approximation risk; This represents the time interval between adjacent sampling times.

[0060] The formula determines the risk exposure duration of the vehicle in the candidate kinematic segment when it is in a state of low safety margin or high approach risk by multiplying the judgment value corresponding to each sampling time with the time interval and then summing the results.

[0061] in, Determined based on collision time and time interval:

[0062] In the formula, when the first Collision time at each sampling moment Less than the collision time threshold , or the Time interval of each sampling moment Less than the time interval threshold When this occurs, it indicates that the sampling time is in a state of low safety margin or high approximation risk. The value is set to 1; when neither of the above two conditions is met, the value will be set to 1. The value is determined to be 0.

[0063] in, and The values ​​were calculated based on the inter-vehicle distance, relative approximation speed, and vehicle speed at the corresponding sampling time. , These are the pre-set collision time threshold and time distance threshold, which can be determined based on test task requirements, safety margin requirements, or historical test experience. It can be determined based on the number of sampling times contained in the candidate kinematic segment. The larger the value, the longer the vehicle remains in a state of low safety margin or high approach risk within that candidate kinematic segment.

[0064] S1-3: Based on the start and end times of the risk evolution stage, candidate kinematic segments are extracted from historical kinematic test data.

[0065] Candidate kinematic segments are continuous time-series data segments extracted from continuous historical kinematic test data. When extracting segments, the starting time of any risk evolution stage can be used as the starting boundary of the candidate kinematic segment, and the ending time of the recovery stage after the end of that risk evolution stage can be used as the ending boundary of the candidate kinematic segment.

[0066] As an example only, candidate kinematic segments may include one or more of the following: risk approach segment, rapid deceleration response segment, lateral dynamic anomaly segment, recovery control anomaly segment, and segment to be distinguished for study. For risk approach segment, the time range from when the distance between vehicles begins to continuously shorten to when the risk is mitigated can be selected; for recovery control anomaly segment, the time range from when the risk is eliminated to when the vehicle's motion state returns to stability can be selected.

[0067] By using the above method, continuous segments containing the complete risk evolution process can be preserved from long-term historical kinematic test data, avoiding the loss of the temporal relationship between risk formation, vehicle response and recovery process by only extracting a certain instantaneous outlier.

[0068] S2: Determine perturbation parameters and kinematic evaluation indicators based on candidate kinematic segments, construct candidate evaluation methods based on candidate kinematic segments, perturbation parameters, and kinematic evaluation indicators, and generate multiple perturbation test segments based on multiple perturbation parameter values ​​of perturbation parameters.

[0069] Candidate evaluation methods are combinations of candidate kinematic segments, perturbation parameters, and kinematic evaluation indices, used to define which historical segment to use as the basis, which test conditions to adjust, and which evaluation results to use for comparison.

[0070] S2 includes: S2-1: Arrange the historical kinematic test data and derived kinematic features in the candidate kinematic segments in chronological order to obtain the candidate kinematic time series data.

[0071] In this embodiment, the candidate kinematic time series data can be represented using a structured time series table. Each row in the structured time series table corresponds to a sampling time, and each column corresponds to time, vehicle speed, vehicle longitudinal acceleration, vehicle lateral acceleration, vehicle-to-vehicle distance, and the corresponding derived kinematic features.

[0072] Before inputting the data into a large model, candidate kinematic time series data can be normalized, segmented, or labeled with fields. Field labeling enables the large model to identify the physical meaning of each value; segmented summarization preserves key extrema, trends, and stage boundaries when candidate kinematic segments are long.

[0073] S2-2: Input the candidate kinematic time series data into the large model, and use the large model to perform semantic analysis on the risk evolution process and vehicle control process in the candidate kinematic segments to obtain a behavioral risk evolution description that includes the risk evolution stage and the key kinematic change features corresponding to each risk evolution stage.

[0074] Behavioral risk evolution description is a structured representation of changes in the explicit behavior of vehicles in candidate kinematic time series data.

[0075] In this embodiment, the behavioral risk evolution description may include one or more of the following: segment type, risk approach phase, control response phase, risk mitigation phase, recovery phase, key kinematic change characteristics, risk escalation triggering factors, control style characteristics, and suggested perturbation direction.

[0076] In this embodiment, the large model can be an instruction-based fine-tuned large language model based on a transformer-decoder network structure. As an example only, the Qwen2.5-7B-Instruct model can be used as the base model, or other large language models with temporal relationship analysis and structured text generation capabilities can be used; this embodiment does not impose specific limitations. The transformer-decoder network analyzes the correlation between different sampling times and different kinematic parameters in the input data through a self-attention mechanism.

[0077] Specifically, according to the sampling time, the vehicle speed, longitudinal acceleration, lateral acceleration, inter-vehicle distance, and corresponding derived kinematic features in the candidate kinematic segments can be arranged into structured input data.

[0078] Among them, derived kinematic features may include one or more of the following: longitudinal acceleration rate of change, lateral acceleration rate of change, distance rate of change, relative approach velocity, collision time, time distance, and risk exposure duration.

[0079] Furthermore, the risk evolution stages and their start and end times determined in S1-2 can be used as input constraints to enable the large model to determine the key kinematic change characteristics corresponding to each risk evolution stage based on the existing stage division.

[0080] In one implementation, historical candidate kinematic time series data and their manually verified behavioral risk evolution descriptions can be used to form training samples, and the pre-trained large language model can be supervised and fine-tuned.

[0081] As an example, 1000 training samples can be selected and divided into training, validation, and test sets in an 8:1:1 ratio. That is, the training set includes 800 training samples, the validation set includes 100 training samples, and the test set includes 100 training samples. During training, the number of training samples input per iteration can be set to 16, and the initial learning rate can be set to... The maximum number of training rounds is set to 5, and the model is validated using a validation set after each round of training.

[0082] The validation criteria can include the results of risk evolution stage identification, the direction of key kinematic changes, and the completeness of output fields. Training can be stopped for the current round if the validation set loss fails to decrease for two consecutive rounds to avoid overtraining.

[0083] As an example only, the preset verification requirements can be a risk evolution stage identification accuracy of no less than 90%, a key kinematic change direction consistency rate of no less than 95%, and an output field completeness rate of no less than 95%. The specific sample size, division ratio, and training parameters can be adjusted according to the amount of available historical test data and the scale of the large model; this embodiment does not impose any limitations on these.

[0084] For training samples with incorrect stage identification, incorrect direction of change, or missing output fields, engineers can correct the behavioral risk evolution description based on the corresponding candidate kinematic time series data, and then re-add the corrected training samples to the training set. In the next round of training, supervised fine-tuning continues using the original training samples and the corrected training samples until the model output meets the preset verification requirements. This allows the large model to gradually learn the correspondence between candidate kinematic time series data and behavioral risk evolution descriptions.

[0085] In this step, the large model is used to transform candidate kinematic time-series data into a structured description that can guide the determination of subsequent perturbation parameters, rather than directly determining whether the candidate evaluation methods have high discriminative power. The version discriminative power of the candidate evaluation methods and the index sensitivity of the kinematic evaluation indicators are still calculated based on the kinematic evaluation indicator results obtained in subsequent steps.

[0086] S2-3: Determine perturbation parameters based on the behavioral risk evolution description, determine kinematic evaluation indicators based on candidate kinematic segments and their derived kinematic features, and construct candidate evaluation methods based on candidate kinematic segments, perturbation parameters, and kinematic evaluation indicators; adjust candidate kinematic segments according to multiple perturbation parameter values ​​to generate multiple perturbation test segments.

[0087] S2-3 includes: S2-3-1: Determine the risk evolution stage to be disturbed based on the behavioral risk evolution description, and determine the disturbance parameters corresponding to the risk evolution stage to be disturbed based on the preset mapping relationship between the risk evolution stage and the disturbance parameters.

[0088] The mapping relationship between risk evolution stages and disturbance parameters can be pre-stored in a mapping table.

[0089] In this embodiment, the perturbation parameters may include one or more of the following: initial distance, relative approach velocity, longitudinal acceleration, lateral acceleration, duration of risk exposure, and recovery phase time window.

[0090]

[0091] The mapping relationship described above is for illustrative purposes only. The same risk evolution stage can correspond to one or more perturbation parameters, and the same perturbation parameter can be used for different risk evolution stages. In practical applications, the appropriate perturbation parameter can be selected based on the data fields contained in the test task and the candidate kinematic segments.

[0092] S2-3-2: Determine the perturbation direction of the perturbation parameters based on the key kinematic change characteristics corresponding to the risk evolution stage to be perturbed in the behavioral risk evolution description.

[0093] The perturbation direction is used to determine whether the perturbation parameter increases or decreases relative to the original parameter value in the candidate kinematic segment.

[0094] For example, when the behavioral risk evolution description indicates that the safety margin compression in the risk approach phase is not significant, the initial distance can be reduced or the relative approach speed can be increased; when the behavioral risk evolution description indicates that the recovery phase duration is short and it is difficult to observe recovery differences, the recovery phase time window can be extended.

[0095] S2-3-3: Under the constraints of vehicle kinematic realizability and kinematic temporal continuity, the disturbance range of the disturbance parameters is determined according to the disturbance parameters and the disturbance direction, and a constrained counterfactual disturbance template is generated based on the disturbance parameters, the disturbance direction and the disturbance range.

[0096] Vehicle kinematics realizability constraints are used to limit the vehicle speed, inter-vehicle distance, longitudinal acceleration, and lateral acceleration after disturbance to within the range that the vehicle or test system can perform.

[0097] The range can be determined based on the vehicle performance boundaries, the capabilities of the testing equipment, or the range of values ​​from historically valid test data.

[0098] Kinematic temporal continuity constraints are used to limit the occurrence of abrupt changes between adjacent sampling times after a disturbance that do not conform to the vehicle's motion patterns.

[0099] For example, when adjusting the longitudinal acceleration within a certain time range, a transition interval can be set at the start and end of that time range, and linear interpolation, piecewise polynomial interpolation, or a smoothing function can be used to make the adjusted longitudinal acceleration continuously connected with the preceding and following time points.

[0100] Furthermore, the constrained counterfactual perturbation template can also be configured with threshold boundary constraints or test validity constraints. Threshold boundary constraints are used to prevent perturbation parameters from exceeding preset test boundaries; test validity constraints are used to exclude parameter combinations that the test system does not allow to be executed.

[0101] As examples only, for delayed-response car-following risk segments, the initial distance can be reduced, the relative approach speed increased, or the risk exposure duration prolonged; for rapid deceleration response segments, the risk trigger time can be advanced, the peak longitudinal deceleration increased, or the risk exposure duration prolonged; for lateral disturbance segments, the peak lateral acceleration increased, the peak duration prolonged, or the recovery phase time window shortened. The above adjustment ratios or time ranges are determined based on vehicle kinematic feasibility constraints.

[0102] S2-3-4: Based on the multiple perturbation parameter values ​​in the constrained counterfactual perturbation template, adjust the candidate kinematic segments to generate multiple perturbation test segments.

[0103] Multiple perturbation parameter values ​​are used to form multi-level test conditions for the same perturbation parameter, ranging from weak to strong or from small to large.

[0104] For example, when the perturbation parameter is the initial distance, multiple initial distance values ​​that decrease or increase progressively can be set around the original initial distance in the candidate kinematic segment; when the perturbation parameter is the relative approximation velocity, multiple progressively changing relative approximation velocity values ​​can be set.

[0105] When adjusting candidate kinematic segments, only the parameters specified by the constrained counterfactual disturbance template and their associated data are modified, while other temporal relationships remain unchanged or are adjusted accordingly based on kinematic temporal continuity constraints. This ensures that multiple disturbance test segments share the same underlying driving process and exhibit comparable differences in disturbance parameters.

[0106] S3: Apply the same set of perturbation test segments to at least two test objects respectively, and obtain the kinematic evaluation index results of each test object under each perturbation test segment.

[0107] The test objects can be different algorithm versions, different control strategies, or different parameter configurations. At least two test objects are used to form a comparison under the same test conditions.

[0108] In this embodiment, the same set of disturbance test segments can be imported into the playback test system, the simulation test system, or a combination of both, so that each test object executes each disturbance test segment sequentially. For different test objects, the input data, disturbance parameter values, start and end times, and evaluation index calculation methods of the disturbance test segments remain consistent.

[0109] The playback test system can convert disturbance test segments into playback inputs that the test object can accept; the simulation test system can construct a simulation test process based on vehicle speed, inter-vehicle distance, longitudinal acceleration, lateral acceleration, and timing relationships in the disturbance test segments. This specification does not limit the specific type of playback test system or simulation test system.

[0110] Kinematic evaluation metrics may include one or more of the following: minimum collision time, minimum time distance, maximum longitudinal acceleration, maximum lateral acceleration, maximum longitudinal acceleration rate of change, maximum lateral acceleration rate of change, duration of risk exposure, and recovery time.

[0111] The recovery time, which characterizes the time required for the vehicle's kinematic state to return to a relatively stable state after the risk has been eliminated, can be determined by the following formula:

[0112] In the formula, This indicates the recovery time of the tested object under the corresponding perturbation test segment. This indicates the moment when the tested object first meets the risk removal conditions. This indicates the moment when the tested object first meets the preset stability conditions after the risk is eliminated.

[0113] Among them, the risk elimination conditions can be determined based on the collision time, time distance, or distance between vehicles.

[0114] For example, when the collision time is not less than a preset collision time threshold and the time interval is not less than a preset time interval threshold, the corresponding sampling time is determined as the risk clearance time. .

[0115] Preset stability conditions can be determined based on changes in vehicle speed, longitudinal acceleration, lateral acceleration, or distance between vehicles.

[0116] For example, when the changes in the aforementioned kinematic parameters within a preset continuous time range are all within the corresponding stable threshold range, the sampling time that first satisfies this condition is determined as... .

[0117] The collision time, time distance, vehicle speed, vehicle longitudinal acceleration, vehicle lateral acceleration, and distance between vehicles are obtained by reading or calculating the kinematic test data output by the playback test system or simulation test system after the corresponding disturbance test segment is executed based on the test object.

[0118] This formula determines the recovery time required for a vehicle to return to kinematic stability from the moment the risk is eliminated by calculating the time difference between the moment the vehicle regains stability and the moment the risk is eliminated. The larger the value, the longer it takes for the vehicle to return to a stable state after the risk has been eliminated.

[0119] S4: Based on the differences in kinematic evaluation index results between different test objects under the same perturbation test segment, determine the version discrimination of candidate evaluation methods, and based on the degree of change of kinematic evaluation index results with perturbation parameter values, determine the index sensitivity of kinematic evaluation indexes.

[0120] Version discrimination is used to characterize the ability of a candidate evaluation method to amplify the differences in results between different test subjects under the same set of perturbation test segments. The larger the version discrimination, the easier it is for the combination of the corresponding candidate kinematic segments, perturbation parameters, and kinematic evaluation indicators to reveal the performance differences between different test subjects.

[0121] For any two tested objects, the first The degree of variability of a single kinematic evaluation index is determined according to the following formula:

[0122] In the formula, Indicates the first The degree of difference between individual indicators corresponding to each kinematic evaluation indicator This indicates the number of perturbation test segments. Indicates the first tested object in the th... The first perturbation test segment under the first Results of kinematic evaluation indicators Indicates that the second tested object is in the first... The first perturbation test segment under the first Results of kinematic evaluation indicators Indicates the first The normalized range corresponding to each kinematic evaluation index This indicates a preset positive number.

[0123] The first and second tested objects can be different algorithm versions, different control strategies, or different parameter configurations. After applying the same set of disturbance test segments to the first and second tested objects respectively, the kinematic test data output from the playback test system or simulation test system is calculated... and .

[0124] For example, when the first When the kinematic evaluation index is the minimum collision time, the two tested objects are respectively from the 1st... The minimum value is selected from the collision time series under each perturbation test segment as the corresponding kinematic evaluation index result.

[0125] Normalization range This is used to reduce the differences in numerical units and scales among different kinematic evaluation indicators. According to the The value of a kinematic evaluation index can be determined by the difference between the maximum and minimum values ​​in historical kinematic test data, or it can be determined based on a pre-set valid value range.

[0126] For example, the first one in the historical test data can be... The difference between the maximum and minimum values ​​of each kinematic evaluation index is used as... When the maximum and minimum values ​​are the same, a preset positive number is added. Avoid having a denominator of zero.

[0127] The formula first calculates the first [number] test segment under each identical disturbance test segment for the two tested objects. The absolute differences between the results of each kinematic evaluation index are then normalized using a normalization range, and finally... The average of the normalized differences corresponding to the perturbation test segments is used to obtain the first... The degree of difference between individual indicators corresponding to each kinematic evaluation indicator. The larger the value, the higher the value of the test object under the same set of disturbance test segments. The more significant the difference in the results of a kinematic evaluation index, the more effective that kinematic evaluation index is in distinguishing between two test subjects.

[0128] In S4, the version differentiation is determined based on the individual index differences corresponding to multiple kinematic evaluation indicators, and its calculation logic is as follows:

[0129]

[0130] In the formula, Indicates version differentiation. This indicates the number of kinematic evaluation metrics used in the version discrimination calculation. Indicates the first Each kinematic evaluation indicator has a corresponding discrimination weight, and the sum of all discrimination weights is 1. The discrimination weights can be preset based on the focus of the test task or historical validation results.

[0131] As an example only, when the kinematic evaluation metrics involved in the calculation include minimum collision time, risk exposure duration, and recovery time, the corresponding discrimination weights can be set to 0.4, 0.35, and 0.25, respectively; when the test task focuses more on collision risk, the discrimination weight corresponding to minimum collision time can also be increased.

[0132] Index sensitivity is used to characterize the degree to which the results of kinematic evaluation indices respond to changes in perturbation parameter values.

[0133] If the perturbation parameter value changes but the kinematic evaluation index result remains basically unchanged, then the kinematic evaluation index is difficult to reflect the change in test conditions; if the kinematic evaluation index result changes significantly with the perturbation parameter value, then the kinematic evaluation index is more suitable as the evaluation basis for the candidate evaluation method.

[0134] Determining the sensitivity of any kinematic evaluation index involves: arranging multiple perturbation parameter values ​​in numerical order, and then determining the sensitivity of the kinematic evaluation index based on the response degree of the kinematic evaluation index results corresponding to adjacent perturbation parameter values. The sensitivity is determined according to the following formula:

[0135] In the formula, This indicates the sensitivity of the kinematic evaluation index corresponding to the currently tested object. This indicates the number of perturbation test segments that correspond one-to-one with multiple perturbation parameter values. and These represent the adjacent numbers arranged in numerical order. The first disturbance parameter value and the second One disturbance parameter value, and These represent the current measured object at the [number]th ... The first disturbance parameter value and the second The kinematic evaluation index results obtained under the perturbation test segment corresponding to each perturbation parameter value; This indicates a preset positive number.

[0136] When multiple test objects exist, the kinematic evaluation index results obtained for each test object under multiple perturbation test segments are substituted into the above formula to obtain the index sensitivity corresponding to each test object. This allows for the independent determination of the responsiveness of the kinematic evaluation index results of each test object to changes in perturbation parameter values, without mixing the kinematic evaluation index results across different test objects.

[0137] When the sensitivity of the index corresponding to each tested object reaches the preset index sensitivity threshold, it is determined that the kinematic evaluation index meets the index sensitivity requirement.

[0138] in, The settings are determined based on the number of perturbation parameter values ​​set in the constrained counterfactual perturbation template, and each perturbation parameter value corresponds to the generation of a perturbation test segment.

[0139] and It is obtained by reading multiple perturbation parameter values ​​from the constrained counterfactual perturbation template and sorting them according to their numerical values.

[0140] and By the corresponding first The first perturbation test segment and the first The results are obtained from the kinematic test data output by the playback test system or simulation test system after each perturbation test segment is applied to the same test object.

[0141] It is preset by the system to prevent the denominator from being zero when adjacent disturbance parameter values ​​are the same or the difference is too small.

[0142] This formula calculates the degree of change in the kinematic evaluation index result when the adjacent perturbation parameter value changes by one unit, and averages the calculation results corresponding to each adjacent perturbation parameter value to obtain the index sensitivity of the kinematic evaluation index. The larger the value, the more significant the response of the kinematic evaluation index to changes in the disturbance parameters.

[0143] In this embodiment, version discrimination and index sensitivity evaluate candidate evaluation methods from different perspectives. Version discrimination is used to determine whether there are significant differences between different test subjects under the same test conditions; index sensitivity is used to determine whether the kinematic evaluation index used produces an effective response when the test conditions change step by step. Together, they are used to eliminate candidate evaluation methods that only show large differences occasionally under a certain test condition, or that cannot distinguish between different test subjects although the test conditions change.

[0144] S5: Based on the version discrimination and the sensitivity of the kinematic evaluation index, determine the high discrimination evaluation method from the candidate evaluation methods.

[0145] S5 includes: S5-1: Compare the version discrimination with the preset version discrimination threshold, and compare the index sensitivity of the kinematic evaluation index corresponding to each tested object with the preset index sensitivity threshold.

[0146] The preset version discrimination threshold and preset index sensitivity threshold can be determined based on the statistical distribution of historical candidate evaluation methods, the calculation results of known effective evaluation methods, or the requirements of the test task.

[0147] For example, the preset quantiles of the calculation results corresponding to historical candidate evaluation methods can be used as the corresponding thresholds, or the thresholds can be calibrated based on the results of manual verification.

[0148] S5-2: When the version differentiation reaches the preset version differentiation threshold, the index sensitivity of the kinematic evaluation index corresponding to each tested object reaches the preset index sensitivity threshold, and the corresponding disturbance test segment satisfies the vehicle kinematics realizability constraint and the kinematics temporal continuity constraint, the combination of the corresponding candidate kinematic segment, disturbance parameter and kinematic evaluation index is determined as the target evaluation combination.

[0149] By simultaneously setting version differentiation conditions, index sensitivity conditions, and disturbance test segment constraints, it is possible to retain a target evaluation combination that can distinguish different tested objects, effectively respond to changes in test conditions, and be executed by the test system.

[0150] S5-3: Generate highly discriminative assessment methods based on the target assessment combination.

[0151] In S5-3, the methods for generating high-discrimination assessments include: S5-3-1: Extract the candidate kinematic segments, perturbation parameters, and kinematic evaluation indicators from the target evaluation combination; determine the range of perturbation parameters based on the multiple perturbation parameter values ​​corresponding to the perturbation parameters; and determine the test execution conditions and judgment conditions.

[0152] Test execution conditions may include one or more of the following: the number of objects under test, the execution order of disturbance test segments, the type of test system, initial state requirements, and data recording fields. Judgment conditions may include: version discrimination reaching a preset version discrimination threshold, index sensitivity reaching a preset index sensitivity threshold, or the results of a specified kinematic evaluation index satisfying a preset comparison relationship.

[0153] S5-3-2: Associate candidate kinematic segments, perturbation parameter ranges, test execution conditions, kinematic evaluation indicators, and judgment conditions to generate evaluation method configuration data that can be invoked by at least one of the playback test system and the simulation test system.

[0154] The evaluation method configuration data can be stored in structured files, database records, or sets of interface parameters.

[0155] As an example only, the evaluation method configuration data may include the evaluation method identifier, the data location of the candidate kinematic segments, the name of the perturbation parameter, the range of the perturbation parameter, the value of each perturbation parameter, the test execution conditions, the calculation rules of the kinematic evaluation index, and the judgment conditions.

[0156] When the playback test system or simulation test system calls the evaluation method configuration data, it can read candidate kinematic segments and multiple perturbation parameter values, sequentially generate or load the corresponding perturbation test segments, perform tests on the specified test object, and output test results according to the unified kinematic evaluation index calculation rules. Therefore, the high-discrimination evaluation method can be reused in subsequent algorithm version comparisons, control strategy comparisons, or parameter configuration comparisons.

[0157] The above method will be illustrated with an example below.

[0158] During a car-following braking history test, historical kinematic test data showed that the distance between the vehicle and the vehicle in front continuously decreased, the collision time gradually decreased, and then the longitudinal deceleration of the vehicle increased, restoring the distance between the vehicles. The system determines the risk approach phase, control response phase, risk mitigation phase, and recovery phase based on derived kinematic characteristics, and extracts candidate kinematic segments from the historical kinematic test data.

[0159] The candidate kinematic time-series data corresponding to the candidate kinematic segment were input into the large model. The behavioral risk evolution description output by the large model shows that the distance between vehicles continuously shortens during the risk approach phase, and the start time of longitudinal deceleration in the control response phase is relatively late. Based on the mapping relationship between the risk evolution phase and the disturbance parameter, the initial distance was determined as the disturbance parameter, and the reduction of the initial distance was determined as the disturbance direction.

[0160] Under the constraints of vehicle kinematic realizability and kinematic temporal continuity, multiple progressively decreasing perturbation parameter values ​​are set around the original initial distance to generate the same set of perturbation test segments. These perturbation test segments are then applied to both the old and new algorithm versions, and the minimum collision time for each perturbation test segment is recorded.

[0161] Version discrimination is determined based on the difference in minimum collision time between the two algorithm versions under the same disturbance test segment. The sensitivity of each algorithm version is determined based on the degree to which the minimum collision time changes with the initial distance for both the old and new algorithm versions. When the version discrimination reaches a preset version discrimination threshold, the sensitivity of both algorithm versions reaches a preset sensitivity threshold, and each disturbance test segment satisfies the vehicle kinematics realizability constraint and the kinematic temporal continuity constraint, the combination of the candidate kinematic segment, initial distance, and minimum collision time is determined as the target evaluation combination, and corresponding high-discrimination evaluation method configuration data is generated.

[0162] It should be noted that the following braking process, disturbance parameters, and kinematic evaluation indicators in the above examples are only used to illustrate the implementation method of this embodiment and do not constitute a limitation on the type of candidate kinematic segments, disturbance parameter type, or kinematic evaluation indicator type. For rapid deceleration response, lateral disturbance, or recovery control process, the corresponding candidate kinematic segments, disturbance parameters, and kinematic evaluation indicators can be determined according to the same processing logic.

[0163] It should be noted that the threshold, disturbance range, and weight in this embodiment are used to implement the corresponding filtering, constraint, and weighted calculation. Their specific values ​​can be set according to the vehicle type, test task, test system capability, and historical test results, and do not affect the processing logic of the method described in this embodiment.

[0164] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for automatically generating high-discrimination evaluation scores for intelligent driving based on historical test data, characterized in that: include: S1: Obtain historical kinematic test data and determine candidate kinematic segments based on the historical kinematic test data; S2: Determine perturbation parameters and kinematic evaluation indicators based on the candidate kinematic segments, construct candidate evaluation methods based on the candidate kinematic segments, the perturbation parameters, and the kinematic evaluation indicators, and generate multiple perturbation test segments according to multiple perturbation parameter values ​​of the perturbation parameters; S3: Apply the same set of disturbance test segments to at least two test objects respectively, and obtain the kinematic evaluation index results of each test object under each disturbance test segment; S4: Based on the differences between the kinematic evaluation index results of different test objects under the same perturbation test segment, determine the version discrimination of the candidate evaluation method, and based on the degree of change of the kinematic evaluation index results with the perturbation parameter value, determine the index sensitivity of the kinematic evaluation index. S5: Based on the version discrimination and the index sensitivity of the kinematic evaluation index, determine the high discrimination evaluation method from the candidate evaluation methods.

2. The method for automatically generating high-discrimination intelligent driving evaluations based on historical test data according to claim 1, characterized in that, S1 includes: S1-1: Obtain continuous historical kinematic test data, and construct derived kinematic features based on the historical kinematic test data; S1-2: Determine the risk evolution stage based on the changes in the derived kinematic characteristics over a continuous time range; S1-3: Based on the start and end times of the risk evolution stage, extract the candidate kinematic segments from the historical kinematic test data.

3. The automatic generation method for high-discrimination intelligent driving evaluation based on historical test data according to claim 2, characterized in that, S2 include: S2-1: Arrange the historical kinematic test data and derived kinematic features in the candidate kinematic segments in chronological order to obtain candidate kinematic time series data; S2-2: Input the candidate kinematic time series data into the large model, and use the large model to perform semantic analysis on the risk evolution process and vehicle control process in the candidate kinematic segments to obtain a behavioral risk evolution description including the risk evolution stage and the key kinematic change features corresponding to each risk evolution stage; S2-3: Determine perturbation parameters based on the behavioral risk evolution description, determine kinematic evaluation indicators based on the candidate kinematic segments and their derived kinematic features, and construct the candidate evaluation method based on the candidate kinematic segments, the perturbation parameters, and the kinematic evaluation indicators; adjust the candidate kinematic segments according to multiple perturbation parameter values ​​of the perturbation parameters to generate the multiple perturbation test segments.

4. The method for automatically generating high-discrimination intelligent driving evaluations based on historical test data according to claim 3, characterized in that, S2-3 includes: S2-3-1: Determine the risk evolution stage to be disturbed based on the behavioral risk evolution description, and determine the disturbance parameters corresponding to the risk evolution stage to be disturbed based on the preset mapping relationship between the risk evolution stage and the disturbance parameters; S2-3-2: Determine the perturbation direction of the perturbation parameter based on the key kinematic change characteristics corresponding to the risk evolution stage to be perturbed in the behavioral risk evolution description; S2-3-3: Under the constraints of vehicle kinematic realizability and kinematic temporal continuity, the disturbance range of the disturbance parameter is determined according to the disturbance parameter and the disturbance direction, and a constrained counterfactual disturbance template is generated based on the disturbance parameter, the disturbance direction and the disturbance range; S2-3-4: Adjust the candidate kinematic segments according to the multiple perturbation parameter values ​​in the constrained counterfactual perturbation template to generate the multiple perturbation test segments.

5. The method for automatically generating high-discrimination intelligent driving evaluations based on historical test data according to claim 1, characterized in that, In S4, for any two tested objects, the first... The calculation logic for the single-index difference corresponding to each kinematic evaluation index is as follows: In the formula, Indicates the first The degree of difference between individual indicators corresponding to each kinematic evaluation indicator This indicates the number of perturbation test segments. Indicates the first tested object in the th... The first perturbation test segment under the first Results of kinematic evaluation indicators Indicates that the second tested object is in the first... The first perturbation test segment under the first Results of kinematic evaluation indicators Indicates the first The normalized range corresponding to each kinematic evaluation index This indicates a preset positive number.

6. The method for automatically generating high-discrimination intelligent driving evaluations based on historical test data according to claim 5, characterized in that, In S4, the version discrimination is determined based on the single-index difference corresponding to multiple kinematic evaluation indicators, and its calculation logic is as follows: In the formula, This indicates the version differentiation. This indicates the number of kinematic evaluation metrics used in the version differentiation calculation. Indicates the first The discrimination weights corresponding to each kinematic evaluation index.

7. The method for automatically generating high-discrimination intelligent driving evaluations based on historical test data according to claim 6, characterized in that, In S4, determining the sensitivity of any kinematic evaluation index for any tested object includes: arranging multiple perturbation parameter values ​​in numerical order according to their values, and determining the sensitivity of the kinematic evaluation index based on the response degree of the kinematic evaluation index results corresponding to adjacent perturbation parameter values. The calculation logic for the sensitivity of the kinematic evaluation index is as follows: In the formula, This indicates the sensitivity of the kinematic evaluation index. This indicates the number of perturbation test segments that correspond one-to-one with the multiple perturbation parameter values. and They represent the adjacent numbers arranged in the numerical order. The first disturbance parameter value and the second One disturbance parameter value, and These respectively represent the tested object at the th The first disturbance parameter value and the second The kinematic evaluation index results obtained under the perturbation test segment corresponding to each perturbation parameter value This indicates a preset positive number.

8. The method for automatically generating high-discrimination intelligent driving evaluations based on historical test data according to claim 7, characterized in that, S5 include: S5-1: Compare the version discrimination with the preset version discrimination threshold, and compare the index sensitivity of the kinematic evaluation index corresponding to each tested object with the preset index sensitivity threshold respectively; S5-2: When the version differentiation reaches the preset version differentiation threshold, the index sensitivity of the kinematic evaluation index corresponding to each tested object reaches the preset index sensitivity threshold, and the corresponding disturbance test segment satisfies the vehicle kinematics realizability constraint and the kinematics temporal continuity constraint, the combination of the corresponding candidate kinematic segment, disturbance parameter and the kinematic evaluation index is determined as the target evaluation combination. S5-3: Generate the high-discrimination assessment method based on the target assessment combination.

9. The method for automatically generating high-discrimination intelligent driving evaluations based on historical test data according to claim 8, characterized in that, In S5-3, generating the high-discrimination evaluation method includes: S5-3-1: Extract the candidate kinematic segments, the perturbation parameters, and the kinematic evaluation indicators from the target evaluation combination; determine the range of perturbation parameters based on the multiple perturbation parameter values ​​corresponding to the perturbation parameters; and determine the test execution conditions and judgment conditions. S5-3-2: Associate the candidate kinematic segments, the perturbation parameter range, the test execution conditions, the kinematic evaluation index, and the judgment conditions to generate evaluation method configuration data that can be called by at least one of the playback test system and the simulation test system.

Images