Automatic driving personification evaluation method based on real driver behavior modeling

Abstract

The invention relates to the technical field of automobile automatic driving, in particular to an automatic driving personification evaluation method based on driver behavior modeling. Comprising the following steps: S1, driving scene extraction: acquiring a multi-source data set from a real driver, and extracting driving scene classification according to the multi-source data set; s2, clustering driving styles: aiming at driving behavior data of a mobile phone driver in various driving scenes, identifying and dividing behavior styles of the driver through a clustering algorithm; s3, aiming at each driving scene and style category, selecting a car-following and lane-changing behavior model, calibrating model parameters by utilizing real driver natural driving data of a corresponding style, and establishing a driver reference model of multiple driving scenes and multiple driving styles; and S4, evaluation index system construction: constructing a key performance index set based on the calibrated driver reference model output characteristics and the automatic driving system test requirements. And S5, anthropomorphic evaluation realization: obtaining automatic driving actual measurement data, inputting the automatic driving actual measurement data into the driver reference model, obtaining a prediction behavior sequence output by the driver reference model, and outputting an anthropomorphic comprehensive evaluation result.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and specifically to an anthropomorphic evaluation method for automated driving based on driver behavior modeling. Background Technology

[0002] Intelligent connected vehicles are equipped with advanced onboard sensors, controllers, actuators, and other devices, integrating modern communication and network technologies to achieve intelligent information exchange and sharing between vehicles and people, vehicles, roads, and the cloud. This enables vehicles to possess environmental perception, intelligent decision-making, and collaborative control functions, achieving comprehensive safety, energy efficiency, environmental friendliness, and comfortable driving, gradually replacing human operation. Intelligent connected vehicles are a crucial direction for the transformation and upgrading of my country's automotive industry, and are now approaching the critical stage of mass production and road deployment. As a vital part of evaluating the safety performance of autonomous driving systems, the autonomous driving testing and evaluation system helps researchers understand system problems and determine optimization directions. Therefore, autonomous driving vehicle evaluation methods play a crucial role in the development of intelligent connected vehicles.

[0003] With the rapid development of autonomous driving technology, the scientific and objective evaluation of the degree of anthropomorphism in driving behavior has become a key technical requirement. Currently, commonly used anthropomorphism evaluation methods in the industry mainly rely on autonomous driving data obtained from simulations or real-vehicle tests. Based on pre-set experience thresholds or existing research results, key performance indicators are quantitatively scored, and a comprehensive scoring system is constructed through weight allocation. While these methods achieve quantitative evaluation to a certain extent, their evaluation benchmarks are mostly static experience values, lacking a reflection of the dynamic characteristics of real driver behavior. This results in insufficient consistency between the evaluation results and actual human driving experience, making it difficult to comprehensively and accurately assess the anthropomorphism level of autonomous driving systems. Existing technologies also include some research attempting to introduce driver behavior models, but most models are still based on theoretical assumptions or simplified scenario construction, failing to fully integrate real driving data from multiple scenarios. This leads to deviations between model outputs and real driving behavior, limiting the accuracy and reliability of the evaluation.

[0004] Therefore, there is an urgent need for an anthropomorphic evaluation method for autonomous driving that can more realistically reflect driving behavior characteristics, has scene adaptability, and has strong interpretability of the evaluation process, so as to make up for the shortcomings of existing technologies in terms of the realism of behavior modeling and the consistency of evaluation. Summary of the Invention

[0005] The technical problem solved by this invention is to provide an automated driving anthropomorphic evaluation method based on driver behavior modeling, which can directly compare the multi-scenario driver model with the measured true value, significantly improving the accuracy and authenticity of the evaluation, and effectively overcoming the limitations of traditional methods in terms of the dynamism of behavioral representation and the objectivity of the benchmark.

[0006] The basic solution provided by this invention is an automated driving anthropomorphic evaluation method based on driver behavior modeling, comprising the following steps: S1. Driving Scene Extraction: Obtain multi-source datasets from real drivers. The multi-source datasets include at least aerial photography datasets and in-vehicle natural driving datasets. Extract driving scene classifications based on the multi-source datasets. The driving scenes include at least stable following, congested following, and lane changing. S2. Driving Style Clustering: Based on driving behavior data of drivers in various driving scenarios, a clustering algorithm is used to identify and classify the drivers' behavior styles, including aggressive, conservative, and normal driving styles. S3. For each driving scenario and style category, select the following and lane-changing behavior models, use the natural driving data of real drivers of the corresponding style to calibrate the model parameters, and establish a driver benchmark model for multiple driving scenarios and multiple driving styles. S4. Evaluation Index System Construction: Based on the output characteristics of the calibrated driver benchmark model and the testing requirements of the autonomous driving system, a set of key performance indicators is constructed from three dimensions: safety, comfort, and efficiency. S5. Human-like evaluation implementation: Obtain actual test data of autonomous driving, input the actual test data of autonomous driving into the driver benchmark model, obtain the output predicted behavior sequence, use KS test to check the consistency between the distribution of actual data and the distribution of predicted data, and combine the threshold criteria of key performance indicators to construct a hybrid evaluation criterion and output human-like comprehensive evaluation results.

[0007] The principle and advantages of this invention are as follows: Typical driving scenarios are extracted, such as stable following, congested following, and lane changing. Then, human driving behavior data in these scenarios is style-classified, for example, identifying aggressive drivers who drive quickly and conservative drivers who drive steadily. Next, real-world data with these style classifications is used to calibrate mathematical models of following and lane-changing behaviors, enabling these models to drive like a real person. Then, a set of evaluation indicators is summarized from aspects such as safety, comfort, and efficiency. Finally, data from actual autonomous vehicle operation is input into the previously built models that mimic human driving to see if the behavior of the two is similar, and combined with whether specific indicators are met, a comprehensive and quantitative anthropomorphic evaluation score is given.

[0008] Compared to existing technologies, the advantage of this solution lies in making the evaluation criteria derived from real drivers, thus being more objective and closer to actual driving experience. It avoids subjective biases caused by artificially set standards, making the judgment on whether the autonomous driving system's driving resembles that of a real driver more convincing. Through scenario- and style-specific modeling, this invention can flexibly adapt to different traffic conditions and driving habits, resulting in more detailed and comprehensive evaluation results. It not only provides an overall score but also helps engineers specifically analyze where the autonomous driving system's driving is not quite human-like enough, thereby enabling targeted improvements.

[0009] Furthermore, S2 includes the following steps: S21. Use the K-means algorithm to perform unsupervised clustering of driving behavior data: S22. For stable car-keeping and congested car-keeping driving scenarios, select the maximum value of the reciprocal of TTC. The average THW across all times in each following event. and acceleration variance As a clustering feature variable; S23. For lane-changing scenarios, select the maximum lateral acceleration. And lane change segment length LC as a clustering feature variable; S24. Perform K-Means clustering on the datasets of each driving scenario. Calculate the silhouette coefficients corresponding to different numbers of clusters k to evaluate the clustering quality, quantify the density and classification degree of the clustering effect, and select the k value with the optimal silhouette coefficient as the best number of driving style categories to complete the division and label definition of driving styles. In particular, the threshold capping method is used to process the feature variable data before clustering to suppress noise interference.

[0010] For the extracted following and lane-changing scenarios, several key behavioral feature indicators are selected. For example, in following scenarios, the "maximum value of the reciprocal of the collision time" reflects the driver's maximum tolerable risk level; the "average headway" indicates whether the driver typically follows close or far behind other vehicles; and the "magnitude of acceleration fluctuation" measures whether the driver's driving is smooth or abrupt. In lane-changing scenarios, the "maximum value of lateral acceleration" indicates whether the lane-changing action is rapid, and the "lane-changing process duration" indicates whether the action is decisive or sluggish. After selecting the indicators, algorithms such as K-Means are used to automatically group data points into several clusters based on their proximity in the feature space. To ensure accurate and effective classification, the method also introduces a silhouette coefficient to objectively evaluate the clustering effect and uses a threshold capping method to pre-process abnormal noise points in the data to prevent individual extreme data from skewing the entire classification. The advantage is that it achieves automation and objectivity in driving style classification. It no longer relies on subjective human feelings to define what is aggressive or conservative, but instead uses mathematical methods to find the inherent group differences. This data-driven style classification provides data support for building accurate behavioral models for different styles, enabling the final evaluation benchmark to cover a wider and more realistic spectrum of human driving behavior, thus improving the precision and reliability of the entire evaluation system.

[0011] Furthermore, S3 includes the following steps: S31. For stable car-following and congested car-following scenarios, the Intelligent Driving Model (IDM) is selected as the calibration object. The acceleration calculation formula of the IDM model is as follows:

[0012] Where 'a' represents acceleration. Indicates the maximum acceleration. Indicates the current acceleration. Indicates the desired speed. The acceleration index is represented by s, which represents the actual distance to the object being pulled. Indicates the speed difference with the vehicle in front. Indicates the required safe distance. , The preset minimum safe distance is T, the desired headway is b, and the comfort deceleration is b. The parameters to be calibrated include... Six parameters; S232. Using the root mean square percentage error of vehicle speed and distance measurement as the optimization objective of the calibration process, the following objective function is constructed:

[0013] in The smaller the value of the objective function, the smaller the error. It is the model output value of the velocity at time t. It is the true value of the velocity at time t. The model output value is the position at time t. The position is the true value at time t, and N is the total number of observations; S32. Based on the genetic algorithm for automatic calibration, with the goal of minimizing the root mean square percentage error between the measured vehicle trajectory data and the simulation output data of the IDM model, the genetic algorithm is used to globally optimize the six parameters to be calibrated in the IDM model to obtain the optimal parameter combination for the actual driving behavior in the rainy scenario.

[0014] The Intelligent Driving Model (IDM) was chosen as the basic framework. This model can simulate different styles of following behavior by adjusting several key parameters, such as whether the following is close or loose, and whether acceleration is sharp or gentle. The core steps are: first, clarify the mathematical formula of the IDM model, which describes how the vehicle's acceleration is calculated based on its own speed, distance from the vehicle in front, and speed difference. The formula contains six parameters to be determined, such as "expected headway" and "maximum acceleration," and different combinations of these parameters correspond to different driving styles. Then, an optimization approach is proposed to find the most suitable values ​​for these parameters: establishing an objective function to calculate the overall error between the vehicle trajectory (including speed and position) simulated by the model and the actually recorded vehicle trajectory. Next, a global search method, such as a genetic algorithm, is used to automatically and repeatedly try various parameter combinations within a reasonable parameter range, aiming to find the set of parameters that minimizes the aforementioned error. The advantage of this method is that it automates and optimizes model parameter calibration, avoiding the inefficiency and inaccuracy of manually adjusting parameters, and efficiently learning the model parameters that best reproduce human driving behavior from real data. Through this process, a unique IDM model is calibrated for each driving style (such as aggressive following and conservative following). These models constitute the benchmark for human drivers to be compared when evaluating autonomous driving in the future.

[0015] Furthermore, S3 also includes the following steps: S33. For lane-changing scenarios, the Gipps lane-changing model is selected as the calibration object. An objective function is established and an optimization algorithm is used to automatically optimize the model parameters to obtain model parameters that reflect the characteristics of real driver lane-changing behavior.

[0016] Furthermore, S4 includes the following steps: S41. Construction of Evaluation Dimensions and Indicator Types: Based on the output characteristics of the calibrated driver benchmark model and the human-like testing requirements of the autonomous driving system, key indicator types are constructed from three dimensions: safety, comfort, and efficiency. Among them, the indicator types for the safety dimension include collision probability, collision avoidance characteristics, and lane keeping characteristics; the indicator types for the comfort dimension include longitudinal stability, lateral stability, braking smoothness, and starting smoothness; and the indicator types for the efficiency dimension include speed and lane changing efficiency. S42. Specific Quantitative Indicator Definition and System Formation: The various indicator types are refined into key performance evaluation indicators that can be directly measured or calculated, forming a personified evaluation indicator system; wherein, the key performance evaluation indicators include at least: collision time TTC, headway THW, collision avoidance deceleration DRAC, braking distance BD, maximum lateral deviation, longitudinal acceleration, longitudinal acceleration rate of change, lateral acceleration, lateral acceleration rate of change, braking deceleration, starting acceleration, vehicle speed, lane change time and merging time.

[0017] First, several key indicator types are defined for each dimension. For example, in the safety dimension, "probability of collision," "emergency avoidance characteristics," and "lane-keeping ability" are considered; in the comfort dimension, the longitudinal and lateral stability of the vehicle, as well as the smoothness of stopping and starting, are considered; and in the efficiency dimension, the efficiency of driving speed and lane-changing operations is considered. Then, these types are further refined into specific indicators that can be directly calculated from vehicle data. For example, "Time to Collision (TTC)" is used to quantify collision risk, "longitudinal acceleration" and its rate of change are used to quantify ride roughness, and "lane-changing time" is used to quantify the speed of lane-changing maneuvers. The advantage is that a clearly structured and comprehensive evaluation framework is established. Through more than a dozen specific technical indicators, the evaluation process is transformed from subjective judgment to objective measurement.

[0018] Furthermore, S5 includes the following steps: S51. Input the real-time acquired autonomous driving test data into the driver baseline model established by value S3, and obtain the behavior prediction sequence under the same initial conditions and scenarios; S52. Perform a consistency quantification analysis on the measured data sequence of the autonomous driving system and the model prediction data sequence using the KS test. S52 includes the following steps: S521. Calculate the empirical cumulative distribution functions of the model-predicted data sequence, the measured data sequence, and the predicted data sequence, respectively. and ; S522. Calculate the KS statistic D, whereby the KS statistic D is defined as the maximum absolute vertical distance between two empirical cumulative distribution functions:

[0019] in, and These are the empirical cumulative distribution functions of two samples: the measured data sequence and the model prediction data sequence of autonomous driving. n and m are the sizes of the two sample sets, with n=m. S523. Calculate the p-value of the hypothesis test based on the KS statistic D and the sample size, and compare it with the set significance level α.

[0020] Real-time data collected by autonomous vehicles on actual roads or during testing is input into pre-calibrated human driving models representing different driving styles. These models, based on the same starting conditions and scenario assumptions, output predicted behavioral sequences of how a human driver would drive, such as predicted speed change curves and position trajectories. This essentially provides a human reference frame for autonomous driving behavior. Secondly, the KS test is introduced for in-depth analysis. Its principle is not simply comparing values ​​at a specific moment, but rather examining the shape and pattern of the overall data distribution to determine if the two are from the same source. Specifically, the empirical cumulative distribution function curves of the measured data sequence and the model-predicted data sequence are plotted separately. These curves reflect the overall statistical characteristics of the data, such as the range in which most data is concentrated. Then, the maximum vertical distance between the two curves, i.e., the KS statistic D, is calculated, and a p-value is calculated based on the D value and the data volume. If the p-value is greater than a set threshold (e.g., 0.05), the hypothesis that the two distributions are consistent is accepted, and it is considered that the behavior patterns of autonomous driving are not significantly different from human driving in terms of overall statistical regularity. This transforms abstract behavioral similarity evaluation into a rigorous and repeatable statistical testing process. By using the KS test to assess the overall pattern of probability distribution, it captures the dynamic statistical characteristics of behavioral patterns, rather than simply checking whether a few isolated indicators exceed thresholds. This is more robust and comprehensive than simply comparing averages or maximum values, effectively identifying autonomous driving behaviors that may meet key indicators but whose overall behavioral rhythm and distribution patterns deviate from human habits, making the evaluation more in-depth and reliable.

[0021] Furthermore, S5 also includes the following steps: S53. Integrate the statistical results of the KS test with the key performance indicator threshold criteria to construct a hybrid evaluation criterion, calculate the anthropomorphic comprehensive evaluation score, and classify the grades. S53 includes the following steps: S531. Calculate the distribution consistency score based on the p-value of the KS test:

[0022] S532. Based on the set of key performance indicators constructed by S4, calculate the degree to which the measured value of each indicator exceeds the corresponding threshold, and calculate the key indicator score accordingly:

[0023] in ; S533. The distribution consistency score and the key indicator score are weighted and fused to obtain a comprehensive anthropomorphic score. Based on the numerical range of the comprehensive anthropomorphic score, the corresponding anthropomorphic evaluation level is output.

[0024] Furthermore, in S1, the extraction rules for the driving scenario include: (1) Stable following scenario: The extraction conditions are that the distance between the front and rear vehicles is between 7 m and 150 m, the lateral distance difference is less than 0.85 m, the headway is less than 3 s, and the following behavior lasts for more than 8 s. (2) Congested following scenario: The extraction conditions are that the distance between the vehicles in front and behind is less than 7 m, the lateral distance difference is less than 0.85 m, and the following behavior lasts for more than 8 seconds; (3) Lane changing scenario: The extraction condition is that the vehicle changes lanes, and the lateral distance change rate of the vehicle relative to the lane centerline is used as the judgment criterion: Lane change starting point: When the lateral distance change rate is greater than 0.05 and begins to change continuously; Lane change endpoint: When the rate of change of lateral distance is less than 0.05 and the lane returns to a steady state. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating an embodiment of an anthropomorphic evaluation method for autonomous driving based on real driver behavior modeling according to the present invention. Figure 2 This is a schematic diagram of a stable following scenario, representing an embodiment of the human-like evaluation method for autonomous driving based on real driver behavior modeling of the present invention. Figure 3 This is a schematic diagram of a congested following scenario, representing an embodiment of an autonomous driving anthropomorphic evaluation method based on real driver behavior modeling according to the present invention. Figure 4 This is a schematic diagram of a lane-changing scenario, representing an embodiment of the human-like evaluation method for autonomous driving based on real driver behavior modeling of the present invention. Figure 5 This is a distribution diagram of parameter values ​​obtained from the calibration of an embodiment of the human-like evaluation method for autonomous driving based on modeling real driver behavior according to the present invention. Figure 6 This is a calibration error distribution diagram of an embodiment of the anthropomorphic evaluation method for autonomous driving based on real driver behavior modeling of the present invention. Detailed Implementation

[0026] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A human-like evaluation method for autonomous driving based on real driver behavior modeling, characterized by the following steps: S1. Driving Scene Extraction: Obtain multi-source datasets from real drivers. The multi-source datasets include at least aerial photography datasets and in-vehicle natural driving datasets. Extract driving scene classifications based on the multi-source datasets. The driving scenes include at least stable following, congested following, and lane changing.

[0027] Specifically, this embodiment uses a publicly available aerial photography trajectory dataset of a highway as the basic data source. This dataset contains key information such as vehicle trajectory data, road condition data, and surrounding environment monitoring data, and examples of its core data fields are shown in Table 1. Preprocessing is performed on the raw data to ensure that the data quality meets the requirements of subsequent analysis. This includes: filling missing values ​​in vehicle operation data using interpolation, deleting dirty data, and smoothing the data to reduce noise.

[0028] Table 1

[0029] Based on the preprocessed data, three types of typical driving scenario segments are extracted according to the following rules: (1) Stable following scene segment (e.g.) Figure 2 Extraction rules (as shown): The distance between the front and rear vehicles is between 7 m and 150 m, the headway is less than 3 s, the lateral distance difference between the two vehicles is less than 0.85 m, and the following state lasts for more than 8 s.

[0030] (2) Scene clips of traffic congestion and car chases (e.g.) Figure 3 Extraction rules (as shown): The distance between the front and rear vehicles is less than 7 m, the lateral distance difference between the two vehicles is less than 0.85 m, and the following state lasts for more than 8 seconds.

[0031] (3) Lane-changing scene segments (such as...) Figure 4 The extraction rule (as shown) is: when a vehicle changes lanes, the rate of change of the vehicle's lateral distance relative to the lane centerline is used as the criterion. ① Lane change starting point: When the rate of change of lateral distance is greater than 0.05 and begins to change continuously; ② Lane change endpoint: When the rate of change of lateral distance is less than 0.05 and the lane returns to a stable state.

[0032] The number of valid scene fragments obtained after extraction according to the above rules is shown in Table 2: Table 2

[0033] S2. Driving Style Clustering: Based on driving behavior data of mobile phone drivers in various driving scenarios, the driving style of drivers is identified and classified by clustering algorithm. The driving style includes aggressive, conservative and normal.

[0034] S2 includes the following steps: S21. Use the K-means algorithm to perform unsupervised clustering of driving behavior data: S22. For stable car-keeping and congested car-keeping driving scenarios, select the maximum value of the reciprocal of TTC. The average THW across all times in each following event. and acceleration variance As a clustering feature variable; S23. For lane-changing scenarios, select the maximum lateral acceleration. And lane change segment length LC as a clustering feature variable; S24. Perform K-Means clustering on the datasets of each driving scenario. Calculate the silhouette coefficients corresponding to different numbers of clusters k to evaluate the clustering quality, quantify the density and classification degree of the clustering effect, and select the k value with the optimal silhouette coefficient as the best number of driving style categories to complete the division and label definition of driving styles. In particular, the threshold capping method is used to process the feature variable data before clustering to suppress noise interference.

[0035] Specifically, driving style clustering was performed on the three types of scenario segments mentioned above. In the Matlab environment, the K-means algorithm was used to perform cluster analysis on the extracted stable car-following, congested car-following, and lane-changing scenario segments. The clustering feature variables for the car-following scenario were selected as follows: , , The clustering feature variables for lane-changing scenarios are selected as follows: And LC.

[0036] To optimize clustering results and suppress noise interference, the following measures are taken in this implementation: (1) The silhouette coefficient is selected as the evaluation index of clustering quality. This coefficient is comprehensively evaluated by calculating the average distance (cohesion) between a single sample point and other points in its own cluster and the average distance (separation) between it and the nearest other point in the cluster. Its value range is [-1, 1]. The closer the value is to 1, the higher the intra-cluster similarity and the better the inter-cluster separation of the clustering results, and the more reasonable the clustering structure.

[0037] (2) To address the issue of outliers (such as abnormal data points caused by sensor noise or extreme behavior) excessively influencing cluster centers, a "threshold capping" method is used for preprocessing. Specifically, reasonable upper and lower thresholds are set based on the physical meaning and data distribution of each feature variable. For data exceeding the threshold range, its value is replaced with the corresponding threshold, thus limiting the data to a reasonable range, ensuring no information is lost while preventing it from excessively affecting the clustering results.

[0038] After the above process, the clustering results are shown in Table 3-5.

[0039] Table 3. K-means-based clustering results of stable car-following driving styles (mean silhouette coefficient 0.4353)

[0040] Table 4. K-means-based clustering results of driving styles under congested conditions (mean silhouette coefficient 0.4184)

[0041] Table 5. K-means-based clustering results of lane-changing driving styles (mean silhouette coefficient 0.6046)

[0042] S3. For each driving scenario and style category, select the following and lane-changing behavior models, use the natural driving data of real drivers of the corresponding style to calibrate the model parameters, and establish a driver benchmark model for multiple driving scenarios and multiple driving styles.

[0043] S31. For stable car-following and congested car-following scenarios, the Intelligent Driving Model (IDM) is selected as the calibration object. The acceleration calculation formula of the IDM model is as follows:

[0044] Where 'a' represents acceleration. Indicates the maximum acceleration. Indicates the current acceleration. Indicates the desired speed. The acceleration index is represented by s, which represents the actual distance to the object being pulled. Indicates the speed difference with the vehicle in front. Indicates the required safe distance. , The preset minimum safe distance is T, the desired headway is b, and the comfort deceleration is b. The parameters to be calibrated include... Six parameters; S232. Using the root mean square percentage error of vehicle speed and distance measurement as the optimization objective of the calibration process, the following objective function is constructed:

[0045] in The smaller the value of the objective function, the smaller the error. It is the model output value of the velocity at time t. It is the true value of the velocity at time t. The model output value is the position at time t. The position is the true value at time t, and N is the total number of observations; S32. Based on the genetic algorithm for automatic calibration, with the goal of minimizing the root mean square percentage error between the measured vehicle trajectory data and the simulation output data of the IDM model, the genetic algorithm is used to globally optimize the six parameters to be calibrated in the IDM model to obtain the optimal parameter combination for the actual driving behavior in the rainy scenario.

[0046] The Intelligent Driving Model (IDM) is calibrated based on the extracted car-following segments. The parameters to be calibrated include: The six parameters and their reference calibration ranges are shown in Table 6. Table 6 Reference Range for IDM Parameter Calibration

[0047] The root mean square percentage error (RMSPE) of vehicle speed and distance was selected as the optimal performance index for model calibration.

[0048] In Matlab, the Genetic Algorithm Toolbox was used to automatically optimize and calibrate the parameters of the IDM model, aiming to minimize the RMSPE between the measured vehicle trajectory data and the simulation output data of the IDM model, and obtain the optimal parameter combination. The first 100 car-following events were used as calibration examples, and the above optimization and calibration were performed independently for each event. The distribution of the calibrated parameter values ​​is shown below. Figure 5 As shown, the data intuitively illustrates the central tendency and dispersion of each parameter's values ​​under different driving events. The corresponding calibration error distribution is as follows: Figure 6 As shown, this reflects the overall fitting accuracy of the model for different driving events.

[0049] S3 further includes the following steps: S33. For lane-changing scenarios, the Gipps lane-changing model is selected as the calibration object. An objective function is established and an optimization algorithm is used to automatically optimize the model parameters to obtain model parameters that reflect the characteristics of real driver lane-changing behavior.

[0050] Specifically, a similar process to the car-following model is used to model and calibrate lane-changing behavior. The Gipps lane-changing model is selected as the base model. By establishing a corresponding objective function and using optimization algorithms (such as genetic algorithms) to find the optimal parameters within a reasonable parameter range, model parameters that accurately reflect the real lane-changing behavior characteristics under various driving styles are calibrated. The calibration results can be used to subsequently build a complete multi-scenario driver benchmark model library.

[0051] S4. Evaluation Index System Construction: Based on the output characteristics of the calibrated driver benchmark model and the testing requirements of the autonomous driving system, a set of key performance indicators is constructed from three dimensions: safety, comfort, and efficiency.

[0052] S4 includes the following steps: S41. Construction of Evaluation Dimensions and Indicator Types: Based on the output characteristics of the calibrated driver benchmark model and the human-like testing requirements of the autonomous driving system, key indicator types are constructed from three dimensions: safety, comfort, and efficiency. Among them, the indicator types for the safety dimension include collision probability, collision avoidance characteristics, and lane keeping characteristics; the indicator types for the comfort dimension include longitudinal stability, lateral stability, braking smoothness, and starting smoothness; and the indicator types for the efficiency dimension include speed and lane changing efficiency. S42. Specific Quantitative Indicator Definition and System Formation: The various indicator types are refined into key performance evaluation indicators that can be directly measured or calculated, forming a personified evaluation indicator system; wherein, the key performance evaluation indicators include at least: collision time TTC, headway THW, collision avoidance deceleration DRAC, braking distance BD, maximum lateral deviation, longitudinal acceleration, longitudinal acceleration rate of change, lateral acceleration, lateral acceleration rate of change, braking deceleration, starting acceleration, vehicle speed, lane change time and merging time.

[0053] Based on the output characteristics of the calibrated driver model and considering the anthropomorphic testing requirements of autonomous driving systems, a hierarchical evaluation index system was constructed from three dimensions: safety, comfort, and efficiency. The safety dimension includes indicators for collision probability, collision avoidance characteristics, and lane keeping characteristics; the comfort dimension includes indicators for lateral and longitudinal stability, braking smoothness, and start-up smoothness; and the efficiency dimension includes indicators for speed and lane-changing efficiency. This index system, the applicable evaluation scenarios for each indicator (referred to as: car-following, lane changing, and congestion), and the core data items required for calculation are shown in Table 7.

[0054] Table 7. Anthropomorphism Evaluation Index System

[0055] S5. Human-like Evaluation Implementation: Acquire real-world autonomous driving test data, input the test data into the driver baseline model, obtain the output predicted behavior sequence, use the KS test to verify the consistency between the distribution of the test data and the distribution of the predicted data, and combine the threshold criteria of key performance indicators to construct a hybrid evaluation criterion, outputting a human-like comprehensive evaluation result. S5 includes the following steps: S51. Input the real-time acquired autonomous driving test data into the driver baseline model established by value S3, and obtain the behavior prediction sequence under the same initial conditions and scenarios; S52. Perform a consistency quantification analysis on the measured data sequence of the autonomous driving system and the model prediction data sequence using the KS test. S52 includes the following steps: First, the KS test is used to quantitatively evaluate the consistency between the measured data sequence and the model prediction data sequence of autonomous driving from the perspective of the overall probability distribution: S521. Calculate the empirical cumulative distribution functions of the model-predicted data sequence, the measured data sequence, and the predicted data sequence, respectively. and ; S522. Calculate the KS statistic D, whereby the KS statistic D is defined as the maximum absolute vertical distance between two empirical cumulative distribution functions:

[0056] in, and These are the empirical cumulative distribution functions of two samples: the measured data sequence and the model prediction data sequence of autonomous driving. n and m are the sizes of the two sample sets, with n=m. S523. Calculate the p-value of the hypothesis test based on the KS statistic D and the sample size, and compare it with the set significance level α.

[0057] Calculate the p-value for hypothesis testing based on the D-value and sample size. If the p-value is greater than the set significance level (e.g., α=0), the null hypothesis that "the two samples come from the same distribution" is accepted, and the autonomous driving behavior is considered to be consistent with the baseline driving behavior predicted by the model in terms of statistical distribution.

[0058] Based on the above, a hybrid evaluation criterion integrating KS statistical test and key performance indicator threshold judgment is constructed to achieve multi-level and robust evaluation of the degree of anthropomorphism. That is, based on the significance results of the KS test and the judgment results of whether each key indicator exceeds the threshold, the final anthropomorphism evaluation level or score is output according to the preset hybrid decision rule.

[0059] S53. Integrate the statistical results of the KS test with the key performance indicator threshold criteria to construct a hybrid evaluation criterion, calculate the anthropomorphic comprehensive evaluation score, and classify the grades. S53 includes the following steps: S531. Calculate the distribution consistency score based on the p-value of the KS test:

[0060] S532. Based on the set of key performance indicators constructed by S4, calculate the degree to which the measured value of each indicator exceeds the corresponding threshold, and calculate the key indicator score accordingly:

[0061] in ; S533. The distribution consistency score and the key indicator score are weighted and fused to obtain a comprehensive anthropomorphic score. Based on the numerical range of the comprehensive anthropomorphic score, the corresponding anthropomorphic evaluation level is output.

[0062] The degree of anthropomorphism is divided into four levels based on the overall score: ① Highly anthropomorphic (Level 1): ; ② Moderate anthropomorphism (Level 2): ; ③Low-level anthropomorphism (Level 3): 50 ; ④ Insufficient anthropomorphism (Level 4): .

[0063] 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 human-like evaluation method for autonomous driving based on real driver behavior modeling, characterized in that: Includes the following steps: S1. Driving Scene Extraction: Obtain multi-source datasets from real drivers. The multi-source datasets include at least aerial photography datasets and in-vehicle natural driving datasets. Extract driving scene classifications based on the multi-source datasets. The driving scenes include at least stable following, congested following, and lane changing. S2. Driving Style Clustering: Based on driving behavior data of drivers in various driving scenarios, a clustering algorithm is used to identify and classify the drivers' behavior styles, including aggressive, conservative, and normal driving styles. S3. For each driving scenario and style category, select the following and lane-changing behavior models, use the natural driving data of real drivers of the corresponding style to calibrate the model parameters, and establish a driver benchmark model for multiple driving scenarios and multiple driving styles. S4. Evaluation Index System Construction: Based on the output characteristics of the calibrated driver benchmark model and the testing requirements of the autonomous driving system, a set of key performance indicators is constructed from three dimensions: safety, comfort, and efficiency. S5. Human-like evaluation implementation: Obtain actual test data of autonomous driving, input the actual test data of autonomous driving into the driver benchmark model, obtain the output predicted behavior sequence, use KS test to check the consistency between the distribution of actual data and the distribution of predicted data, and combine the threshold criteria of key performance indicators to construct a hybrid evaluation criterion and output human-like comprehensive evaluation results.

2. The method for anthropomorphic evaluation of autonomous driving based on real driver behavior modeling as described in claim 1, characterized in that: S2 includes the following steps: S21. Use the K-means algorithm to perform unsupervised clustering of driving behavior data: S22. For stable car-keeping and congested car-keeping driving scenarios, select the maximum value of the reciprocal of TTC. The average THW across all times in each following event. and acceleration variance As a clustering feature variable; S23. For lane-changing scenarios, select the maximum lateral acceleration. And lane change segment length LC as a clustering feature variable; S24. Perform K-Means clustering on the datasets of each driving scenario. Calculate the silhouette coefficients corresponding to different numbers of clusters k to evaluate the clustering quality, quantify the density and classification degree of the clustering effect, and select the k value with the optimal silhouette coefficient as the best number of driving style categories to complete the division and label definition of driving styles. In particular, the threshold capping method is used to process the feature variable data before clustering to suppress noise interference.

3. The method for anthropomorphic evaluation of autonomous driving based on real driver behavior modeling as described in claim 2, characterized in that: S3 includes the following steps: S31. For stable car-following and congested car-following scenarios, the Intelligent Driving Model (IDM) is selected as the calibration object. The acceleration calculation formula of the IDM model is as follows: Where 'a' represents acceleration. Indicates the maximum acceleration. Indicates the current acceleration. Indicates the desired speed. The acceleration index is represented by s, which represents the actual distance to the object being pulled. Indicates the speed difference with the vehicle in front. Indicates the required safe distance. , The preset minimum safe distance is T, the desired headway is b, and the comfort deceleration is b. The parameters to be calibrated include... Six parameters; S232. Using the root mean square percentage error of vehicle speed and distance measurement as the optimization objective of the calibration process, the following objective function is constructed: in The smaller the value of the objective function, the smaller the error. It is the model output value of the velocity at time t. It is the true value of the velocity at time t. The model output value is the position at time t. The position is the true value at time t, and N is the total number of observations; S32. Based on the genetic algorithm for automatic calibration, with the goal of minimizing the root mean square percentage error between the measured vehicle trajectory data and the simulation output data of the IDM model, the genetic algorithm is used to globally optimize the six parameters to be calibrated in the IDM model to obtain the optimal parameter combination for the actual driving behavior in the rainy scenario.

4. The method for anthropomorphic evaluation of autonomous driving based on real driver behavior modeling as described in claim 3, characterized in that: S3 further includes the following steps: S33. For lane-changing scenarios, the Gipps lane-changing model is selected as the calibration object. An objective function is established and an optimization algorithm is used to automatically optimize the model parameters to obtain model parameters that reflect the characteristics of real driver lane-changing behavior.

5. The method for anthropomorphic evaluation of autonomous driving based on real driver behavior modeling as described in claim 4, characterized in that: S4 includes the following steps: S41. Construction of Evaluation Dimensions and Indicator Types: Based on the output characteristics of the calibrated driver benchmark model and the human-like testing requirements of the autonomous driving system, key indicator types are constructed from three dimensions: safety, comfort, and efficiency. Among them, the indicator types for the safety dimension include collision probability, collision avoidance characteristics, and lane keeping characteristics; the indicator types for the comfort dimension include longitudinal stability, lateral stability, braking smoothness, and starting smoothness; and the indicator types for the efficiency dimension include speed and lane changing efficiency. S42. Specific Quantitative Indicator Definition and System Formation: The various indicator types are refined into key performance evaluation indicators that can be directly measured or calculated, forming a personified evaluation indicator system; wherein, the key performance evaluation indicators include at least: collision time TTC, headway THW, collision avoidance deceleration DRAC, braking distance BD, maximum lateral deviation, longitudinal acceleration, longitudinal acceleration rate of change, lateral acceleration, lateral acceleration rate of change, braking deceleration, starting acceleration, vehicle speed, lane change time and merging time.

6. The method for anthropomorphic evaluation of autonomous driving based on real driver behavior modeling as described in claim 5, characterized in that: S5 includes the following steps: S51. Input the real-time acquired autonomous driving test data into the driver baseline model established by value S3, and obtain the behavior prediction sequence under the same initial conditions and scenarios; S52. Perform a consistency quantification analysis on the measured data sequence of the autonomous driving system and the model prediction data sequence using the KS test. S52 includes the following steps: S521. Calculate the empirical cumulative distribution functions of the model-predicted data sequence, the measured data sequence, and the predicted data sequence, respectively. and ; S522. Calculate the KS statistic D, whereby the KS statistic D is defined as the maximum absolute vertical distance between two empirical cumulative distribution functions: in, and These are the empirical cumulative distribution functions of two samples: the measured data sequence and the model prediction data sequence of autonomous driving. n and m are the sizes of the two sample sets, with n=m. S523. Calculate the p-value of the hypothesis test based on the KS statistic D and the sample size, and compare it with the set significance level α.

7. The method for anthropomorphic evaluation of autonomous driving based on real driver behavior modeling as described in claim 6, characterized in that: S5 further includes the following steps: S53. Integrate the statistical results of the KS test with the key performance indicator threshold criteria to construct a hybrid evaluation criterion, calculate the anthropomorphic comprehensive evaluation score, and classify the grades. S53 includes the following steps: S531. Calculate the distribution consistency score based on the p-value of the KS test: S532. Based on the set of key performance indicators constructed by S4, calculate the degree to which the measured value of each indicator exceeds the corresponding threshold, and calculate the key indicator score accordingly: in ; S533. The distribution consistency score and the key indicator score are weighted and fused to obtain a comprehensive anthropomorphic score. Based on the numerical range of the comprehensive anthropomorphic score, the corresponding anthropomorphic evaluation level is output.

8. The method for anthropomorphic evaluation of autonomous driving based on real driver behavior modeling as described in claim 1, characterized in that: In S1, the extraction rules for driving scenarios include: (1) Stable following scenario: The extraction conditions are that the distance between the front and rear vehicles is between 7 m and 150 m, the lateral distance difference is less than 0.85 m, the headway is less than 3 s, and the following behavior lasts for more than 8 s. (2) Congested following scenario: The extraction conditions are that the distance between the vehicles in front and behind is less than 7 m, the lateral distance difference is less than 0.85 m, and the following behavior lasts for more than 8 seconds; (3) Lane changing scenario: The extraction condition is that the vehicle changes lanes, and the lateral distance change rate of the vehicle relative to the lane centerline is used as the judgment criterion: Lane change starting point: When the lateral distance change rate is greater than 0.05 and begins to change continuously; Lane change endpoint: When the rate of change of lateral distance is less than 0.05 and the lane returns to a steady state.

Images