An evaluation method for defensive driving capability of an autonomous driving system

Abstract

The application relates to a defensive driving capability evaluation method of an automatic driving system, which comprises the following steps: simulating typical defensive driving scenes and expert driver driving behavior data; based on the expert driver driving behavior data, a conditional variational autoencoder (CVAE) is trained to construct an active defense reference driver model; in the typical defensive driving scenes, a reference space-time safety envelope region is generated based on the active defense reference driver model, and a tested automatic driving system is tested under the same scene conditions to obtain corresponding test results; based on the test results, according to the inclusion relationship between the space-time trajectory of the tested automatic driving system and the reference space-time safety envelope region, the safety, efficiency and comfort indexes are calculated to obtain the comprehensive defensive driving capability score of the tested automatic driving system, that is, the evaluation result. Compared with the prior art, the defensive driving capability of the automatic driving system can be effectively and comprehensively evaluated.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving testing technology, and in particular to a method for evaluating the defensive driving capabilities of an autonomous driving system. Background Technology

[0002] With the continuous development of intelligent driving technology, people's expectations for autonomous driving systems are gradually increasing, especially their performance in handling complex dynamic environments. Autonomous driving systems face enormous challenges when dealing with various dynamic situations, usually requiring a higher level of decision-making and planning capabilities. To ensure the safe operation of vehicles, safety testing must be conducted on autonomous driving systems.

[0003] Currently, safety testing and evaluation of autonomous driving systems primarily focus on passive and active safety. Passive safety testing examines whether a vehicle can minimize harm to occupants in the event of a traffic accident. Active safety testing focuses on the vehicle's perception and control systems' response to real-world risks, examining whether the vehicle can prevent accidents through interventions such as warnings, automatic braking, and steering assist when such risks occur. However, these safety testing methods struggle to quantify the defensive driving capabilities of human drivers, thus hindering more rigorous testing and evaluation of autonomous driving systems.

[0004] For human drivers, defensive driving refers to a driving style that avoids traffic accidents by anticipating potential risks in a scenario and taking corresponding measures in advance. Drivers with defensive driving capabilities can accurately anticipate potential risks from other road users or road conditions and take reasonable and effective measures in advance to prevent accidents. Integrating defensive driving capabilities into the evaluation of autonomous driving systems can further ensure the safety of autonomous vehicles, ensuring they achieve "preventative safety." However, traditional methods using questionnaires to collect subjective qualitative evaluations of autonomous driving systems' defensive driving capabilities cannot achieve objective quantitative evaluation. Existing technologies design reference driver models to achieve automated safety testing and evaluation, but these models are mostly based on human-calibrated parameters and use analytical methods to generate deterministic emergency avoidance actions. Such methods are not only limited to a single safety reference baseline but also struggle to characterize the heterogeneity of human driver behavior. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for evaluating the defensive driving capability of an autonomous driving system, which can comprehensively and accurately quantify the defensive driving capability of an autonomous driving system.

[0006] The objective of this invention can be achieved through the following technical solution: a method for evaluating the defensive driving capability of an autonomous driving system, comprising the following steps: S1. Simulate and reproduce typical defensive driving scenarios; S2. Collect driving behavior data of expert drivers in typical defensive driving scenarios; S3. Based on expert driver driving behavior data, an active defense reference driver model that can characterize the probability distribution of defensive driving behavior is constructed by training a Conditional Variational Autoencoder (CVAE). S4. In typical defensive driving scenarios, a reference spatiotemporal safety envelope region is generated based on the active defense reference driver model, and the tested autonomous driving system is tested under the same scenario conditions to obtain the corresponding test results. S5. Based on the test results, according to the inclusion relationship between the spatiotemporal trajectory of the tested autonomous driving system and the reference spatiotemporal safety envelope region, the comprehensive score of the defensive driving capability of the tested autonomous driving system is obtained by calculating safety, efficiency and comfort indicators, thus obtaining the evaluation result.

[0007] Furthermore, S1 specifically refers to the simulation and reproduction of typical defensive driving scenarios based on real vehicle field tests or simulation software. The typical defensive driving scenarios are: typical driving scenarios that require defensive driving behaviors to avoid potential accidents or reduce potential risks, based on open-source natural driving data and accident data and selected by experts.

[0008] Further, S2 includes the following steps: S21. Select human expert drivers with defensive driving capabilities; S22. In the simulated typical defensive driving scenario, control of the vehicle is transferred to an expert driver, and driving behavior data of the expert driver is collected. S23. Among the collected driving behavior data, samples that were successfully avoided or effectively reduced due to the proactive defensive driving behaviors of expert drivers are selected as defensive driving behavior data.

[0009] Further, S3 includes the following steps: S31. Construct the CVAE network architecture, including the master encoder, attention module, conditional encoder, and decoder; S32. During the training phase, the CVAE network is trained using defensive driving behavior data and corresponding defensive driving scenario state information, with the goal of minimizing loss, so that the CVAE learns the characteristics of defensive driving trajectories. S33. In the generation phase, the current scene state information is input into the trained CVAE, and defensive driving trajectory is generated through latent space sampling and decoding.

[0010] Furthermore, in the CVAE network architecture constructed in S31, the attention module is used to extract scene feature vectors with high potential risks in the scene and inject them into the conditional encoder. The conditional encoder is used to encode the received scene feature vectors into scene feature latent variables; The master encoder is used to encode the defensive driving trajectory sequence of an expert driver into latent trajectory feature variables; The decoder is used to concatenate scene feature latent variables and trajectory feature latent variables, and map them to Gaussian distribution parameters in the latent space, by sampling latent variables from the latent space. ,Will By fusing with latent variables of scene features, a planned trajectory for an active defense reference driver model is generated.

[0011] Furthermore, the attention module includes a temporal attention module and a target attention module, wherein the temporal attention module is used to extract temporal feature vectors from scene state information; The target attention module is used to extract target interaction feature vectors from scene state information; The scene feature vector is obtained by concatenating the temporal feature vector with the target interaction feature vector.

[0012] Further, S4 includes the following steps: S41. In typical defensive driving scenarios, the active defense reference driver model is tested multiple times to generate multiple reference trajectories, which are then mapped to a joint time-space coordinate system to define the reference spatiotemporal safety envelope region under a pre-set confidence level. ; S42. Under the same scenario conditions, test the autonomous driving system under test and record the corresponding spatiotemporal trajectory data.

[0013] Further, S5 includes the following steps: S51. Calculate the safety score, efficiency score, and comfort score respectively; S52. Map the spatiotemporal trajectory of the tested autonomous driving system to a spatiotemporal coordinate system and determine the relationship between the tested trajectory and the reference safety envelope region. The inclusion relationship; S53. Calculate the comprehensive score of the defensive driving capability of the tested autonomous driving system.

[0014] Furthermore, the formula for calculating the security score is as follows: in, To proactively defend against the average risk magnitude faced by all reference trajectories of the reference driver model in the scenario, The magnitude of the risks faced by the tested autonomous driving system in the scenario. and All were normalized to the range [0,1]. As a correction factor, when The time characterization shows that the safety level of the tested autonomous driving system is on par with the active defense reference driver model; The formula for calculating the efficiency score is as follows: in, The average efficiency index of all reference trajectories of the active defense reference driver model is calculated. The results of efficiency index calculations for the tested autonomous driving system. This is the scaling factor for the limit value; The formula for calculating the comfort score is as follows: in, The average comfort index of all reference trajectories of the active defense reference driver model is calculated. The results of the comfort index calculation for the tested autonomous driving system. For comfort limits.

[0015] Furthermore, the formula for calculating the comprehensive score of the defensive driving capability of the tested autonomous driving system is as follows: in, The overall score is for defensive driving ability. For safety rating, The spatiotemporal trajectory of the tested autonomous driving system. For efficiency scoring, For comfort rating, , and These are the weighting coefficients. and This is the scaling factor. The higher the value, the stronger the defensive driving ability.

[0016] Compared with the prior art, the present invention has the following advantages: This invention, based on simulating typical defensive driving scenarios, trains a Conditional Variational Autoencoder (CVAE) using driving behavior data from expert drivers to construct an active defensive reference driver model that characterizes the probability distribution of defensive driving behaviors. Then, in typical defensive driving scenarios, a reference spatiotemporal safety envelope region is generated based on the active defensive reference driver model, and the tested autonomous driving system is tested under the same scenario conditions to obtain corresponding test results. Finally, based on the test results, and according to the inclusion relationship between the spatiotemporal trajectory of the tested autonomous driving system and the reference spatiotemporal safety envelope region, a comprehensive score for the defensive driving capability of the tested autonomous driving system is obtained by calculating safety, efficiency, and comfort indicators. This allows the system to learn the defensive driving behaviors of expert drivers and, using the reference driver model as a benchmark, comprehensively evaluate the system through the safety envelope region and multi-dimensional indicators, extending the evaluation system from the traditional "deterministic baseline" to a "reference spatiotemporal safety region," thereby effectively and comprehensively evaluating the defensive driving capability of autonomous driving systems.

[0017] This invention selects samples from expert driver driving behavior data that successfully avoid accidents or effectively reduce potential risks due to the expert drivers' proactive defensive driving behaviors. These samples are used as defensive driving behavior data (including defensive driving trajectories and their corresponding scene state information). The defensive driving behavior data is used to train the CVAE (Continuous Vehicle Automotive Assist) with the goal of minimizing losses. This allows the CVAE to learn the characteristics of defensive driving trajectories, thereby constructing a reference driver model with proactive defensive capabilities. The evaluation focus shifts from emergency avoidance capabilities and the degree of accident damage reduction to the ability to proactively reduce risks before danger occurs. This places higher safety requirements on the tested autonomous driving system and helps promote the continuous iteration and optimization of autonomous driving systems in terms of risk perception, decision-making foresight, and behavioral robustness.

[0018] This invention aggregates multiple sampled active defense reference trajectories into a reference spatiotemporal safety envelope region, defining the acceptable behavioral boundaries of defensive driving within the joint temporal and spatial domain. This preserves the reasonable heterogeneity of human drivers in strategy selection while avoiding evaluation bias caused by using a single trajectory or fixed threshold as an absolute standard.

[0019] Based on the reference spatiotemporal safety envelope region, this invention combines multiple evaluation indicators such as safety, efficiency, and comfort to comprehensively evaluate the defensive driving capability of the tested autonomous driving system. It compares the performance of the autonomous driving system with the active defense reference driver model, thereby quantitatively evaluating the defensive driving capability of the autonomous driving system. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of a blind spot scenario in the embodiment; Figure 3 This is a schematic diagram of the CVAE architecture in the embodiment; Figure 4 The reference spatiotemporal safety envelope region and the test results are shown in the example. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0022] Example like Figure 1 As shown, a method for evaluating the defensive driving capability of an autonomous driving system includes the following steps: S1. Simulate and reproduce typical defensive driving scenarios; S2. Collect driving behavior data of expert drivers in typical defensive driving scenarios; S3. Based on expert driver driving behavior data, an active defense reference driver model that can characterize the probability distribution of defensive driving behavior is constructed by training a conditional variational autoencoder (CVAE). S4. In typical defensive driving scenarios, a reference spatiotemporal safety envelope region is generated based on the active defense reference driver model, and the tested autonomous driving system is tested under the same scenario conditions to obtain the corresponding test results. S5. Based on the test results, according to the inclusion relationship between the spatiotemporal trajectory of the tested autonomous driving system and the reference spatiotemporal safety envelope region, the comprehensive score of the defensive driving capability of the tested autonomous driving system is obtained by calculating safety, efficiency and comfort indicators, thus obtaining the evaluation result.

[0023] When simulating and reproducing typical defensive driving scenarios, these scenarios are specifically selected by experts based on open-source natural driving data and accident data, and are typical scenarios that require defensive driving behaviors to avoid potential accidents or reduce potential risks.

[0024] In this embodiment, blind spot scenarios were selected as typical defensive driving scenarios after expert analysis and screening. Blind spot scenarios include... Figure 2 As shown, the car is traveling forward in the rightmost lane. Due to the obstruction of a stationary vehicle in the adjacent lane on the left, a blind spot is created, preventing the driver from directly observing the traffic situation at the intersection to the left. A bicycle is moving perpendicular to the car's direction of travel within this blind spot. If the car does not take proactive preventative measures before reaching the intersection, there is a risk of collision with the bicycle. In this blind spot scenario, the bicycle's movement trigger distance is... Set to 60m, bicycle speed Set to 3 m / s.

[0025] In practical applications, typical defensive driving scenarios can be simulated and reproduced based on real-vehicle field tests or simulation software. This embodiment simulates and reproduces a typical defensive driving scenario in the simulation software Carla.

[0026] The specific process for collecting driving behavior data from expert drivers in typical defensive driving scenarios includes: Step 21: Select a human expert driver with defensive driving capabilities; This embodiment selected 82 human drivers with defensive driving awareness as expert drivers by designing and distributing a defensive driving ability questionnaire. The questionnaire consisted of 10 scenarios, including animated clips of real-world driving scenarios and multiple-choice questions. The real-world driving scenario clips were taken from dashcam videos in actual road environments and were all dangerous situations. Participants had to choose the most appropriate decision for each scenario from the multiple-choice options. Human drivers with defensive driving awareness were selected based on their scores on the 10 questions.

[0027] Step 22: In a simulated typical defensive driving scenario, control of the vehicle is transferred to an expert driver, and driving behavior data of the expert driver is collected. In this embodiment, expert drivers control their own vehicles to complete corresponding driving tasks in a typical defensive driving scenario reproduced by the simulation software Carla, using the Logitech G29 driving simulator, and their driving behavior data is collected simultaneously. Each expert driver conducts three repeated experiments in the blind spot scenario, and a total of 82×3=246 samples are collected.

[0028] Step 23: From the collected driving behavior data, select samples that successfully avoided accidents or effectively reduced potential risks due to the proactive defensive driving behaviors of expert drivers, and use them as defensive driving behavior data.

[0029] In this embodiment, in the blind spot scenario, the moment when the bicycle begins to move is taken as the potential risk point moment, and the moment when the bicycle enters the vehicle's field of vision is taken as the actual risk point moment. The potential risk point moment refers to the moment when the external environmental factors have not yet evolved into a dangerous state but have the possibility of evolving into an accident, and the actual risk point moment refers to the moment when the external environmental factors have evolved into a dangerous state and are about to induce an accident. Then, samples that perform deceleration or steering avoidance actions between the potential risk point moment and the actual risk point moment are selected as defensive driving behavior samples. Finally, a total of 138 defensive driving behavior samples are obtained.

[0030] The specific process for constructing the active defense reference driver model includes: Step 31: Construct the CVAE network architecture, including the main encoder, attention module, conditional encoder, and decoder; Step 32: During the training phase, the CVAE is trained using the defensive driving trajectory of an expert driver and its corresponding scene state information, with the goal of minimizing the loss, so that the CVAE learns the features of the defensive driving trajectory. Step 33: In the generation phase, the current scene state information is input into the trained CVAE, and defensive driving trajectories are generated through latent space sampling and decoding. The generated defensive driving trajectory is then input into the underlying controller, which calculates and outputs control commands to complete the tracking of the defensive driving trajectory.

[0031] In this embodiment, the architecture of CVAE is as follows: Figure 3 As shown, the scene state information is first input into the attention module, which includes a temporal attention module and a target attention module. These modules extract target features and temporal evolution features with high potential risk, concatenating them into a scene feature vector before injecting it into the conditional encoder. The conditional encoder encodes the received scene feature vector into scene feature latent variables. Simultaneously, the main encoder encodes the expert driver's defensive driving trajectory sequence into trajectory feature latent variables. Next, the scene feature latent variables and trajectory feature latent variables are concatenated into a fusion latent variable and mapped to Gaussian distribution parameters in the latent space. Finally, the decoder samples latent variables from the latent space. Furthermore, by integrating latent variables of scene features, a planned trajectory for an active defense reference driver model is generated.

[0032] In this embodiment, based on defensive driving behavior samples, map information, background vehicle status information, and the vehicle's historical trajectory at potential risk points are input to the temporal attention module and the target attention module. The temporal attention module adopts a self-attention mechanism and outputs a temporal feature vector, while the target attention module adopts a cross-attention mechanism and outputs a target interaction feature vector. The scene feature vector is obtained through feature fusion. The future trajectory of the vehicle at potential risk points is input to the main encoder, and the scene feature vector is input to the conditional encoder. The encoded and fused feature vectors are mapped to Gaussian distribution parameters in the latent space. During the training phase, the loss function aims to minimize the trajectory error term and the Kullback-Leibler (KL) divergence term. CVAE is trained for a total of 500 rounds. After generating the defensive driving trajectory, the model predictive control method is used to complete the trajectory tracking.

[0033] Subsequently, in typical defensive driving scenarios, the constructed active defense reference driver model was tested multiple times, generating multiple reference trajectories, which were mapped to a joint time-space coordinate system to delineate the reference spatiotemporal safety envelope region under a pre-set confidence level. Under the same scenario conditions, the autonomous driving system under test is tested, and the corresponding spatiotemporal trajectory data is recorded.

[0034] In this embodiment, the active defense reference driver model was tested 10 times in a blind spot scenario, and a reference spatiotemporal safety envelope region with a 95% confidence level was defined based on the 10 generated reference trajectories. ,like Figure 4 As shown; under the same conditions in the blind spot scenario, the Intelligent Driver Model (IDM) was selected as the autonomous driving system under test for testing.

[0035] When evaluating the defensive driving capabilities of the tested autonomous driving system, a safety score, an efficiency score, and a comfort score are first calculated. Then, the spatiotemporal trajectory of the tested autonomous driving system is mapped to a spatiotemporal coordinate system, and the relationship between the tested trajectory and the reference safety envelope region is determined. The inclusion relationship is determined; finally, the comprehensive score of the defensive driving capability of the tested autonomous driving system is calculated using the following formula: In the formula, The overall score is for defensive driving ability. For safety rating, The spatiotemporal trajectory of the tested autonomous driving system. For efficiency scoring, For comfort rating, , and These are the weighting coefficients. and This is the scaling factor.

[0036] The safety score calculation formula is as follows: In the formula, To proactively defend against the average risk magnitude faced by all reference trajectories of the reference driver model in the scenario, The magnitude of the risks faced by the tested autonomous driving system in the scenario. and All were normalized to the range [0,1]. As a correction factor, when The time characterization shows that the safety level of the tested autonomous driving system is on par with the active defense reference driver model; In this embodiment, the Discrete and Normalized Driving Area (DNDA) is selected as the risk indicator to calculate the risk level. The formula for calculating DNDA is as follows: In the formula, The ideal drivable area. The actual drivable area is calculated as follows: , , ; The formula for calculating the efficiency score is as follows: In the formula, The average efficiency index of all reference trajectories of the active defense reference driver model is calculated. The results of efficiency index calculations for the tested autonomous driving system. This is the scaling factor for the limit value; In this embodiment, transit time is selected as the efficiency indicator. Calculated , , ; The formula for calculating the comfort score is as follows: In the middle of the time, The average comfort index of all reference trajectories of the active defense reference driver model is calculated. The results of the comfort index calculation for the tested autonomous driving system. For comfort limits.

[0037] In this embodiment, the average jerk is selected as the comfort index. Calculated , , ; In this embodiment, take , , , Because the IDM only performs emergency braking when it is relatively close to the bicycle, its trajectory lies outside the reference spacetime safety envelope region, such as... Figure 4 As shown, it satisfies and Based on the situation, the IDM defensive driving capability comprehensive score was calculated. .

[0038] In practical applications, the tested autonomous driving system The higher the score, the stronger its defensive driving capability. In this embodiment, the IDM's overall defensive driving capability score... A lower score indicates weak defensive driving ability, meaning it lacks the capacity to anticipate risks and take defensive driving actions to reduce potential risks in a given scenario.

[0039] In summary, this solution constructs an active defense reference driver model based on expert driving data, and uses the performance of the active defense reference driver model as a benchmark. Through comprehensive scoring using the safety envelope area and multi-dimensional indicators, it can comprehensively and effectively achieve a quantitative assessment of the defensive driving capabilities of the tested autonomous driving system.

Claims

1. A method for evaluating the defensive driving capability of an autonomous driving system, characterized in that, Includes the following steps: S1. Simulate and reproduce typical defensive driving scenarios; S2. Collect driving behavior data of expert drivers in typical defensive driving scenarios; S3. Based on expert driver driving behavior data, an active defense reference driver model that can characterize the probability distribution of defensive driving behavior is constructed by training a conditional variational autoencoder (CVAE). S4. In typical defensive driving scenarios, a reference spatiotemporal safety envelope region is generated based on the active defense reference driver model, and the tested autonomous driving system is tested under the same scenario conditions to obtain the corresponding test results. S5. Based on the test results, according to the inclusion relationship between the spatiotemporal trajectory of the tested autonomous driving system and the reference spatiotemporal safety envelope region, the comprehensive score of the defensive driving capability of the tested autonomous driving system is obtained by calculating safety, efficiency and comfort indicators, thus obtaining the evaluation result.

2. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 1, characterized in that, S1 specifically refers to the simulation and reproduction of typical defensive driving scenarios based on real vehicle field tests or simulation software. The typical defensive driving scenarios are: typical driving scenarios that require defensive driving behaviors to avoid potential accidents or reduce potential risks, based on open-source natural driving data and accident data and selected by experts.

3. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 1, characterized in that, S2 includes the following steps: S21. Select human expert drivers with defensive driving capabilities; S22. In the simulated typical defensive driving scenario, control of the vehicle is transferred to an expert driver, and driving behavior data of the expert driver is collected. S23. Among the collected driving behavior data, samples that were successfully avoided or effectively reduced due to the proactive defensive driving behaviors of expert drivers are selected as defensive driving behavior data.

4. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 3, characterized in that, S3 includes the following steps: S31. Construct the CVAE network architecture, including the master encoder, attention module, conditional encoder, and decoder; S32. During the training phase, the CVAE network is trained using defensive driving behavior data and corresponding defensive driving scenario state information, with the goal of minimizing loss, so that the CVAE learns the characteristics of defensive driving trajectories. S33. In the generation phase, the current scene state information is input into the trained CVAE, and defensive driving trajectory is generated through latent space sampling and decoding.

5. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 4, characterized in that, In the CVAE network architecture constructed by S31, the attention module is used to extract scene feature vectors with high potential risks and inject them into the conditional encoder. The conditional encoder is used to encode the received scene feature vectors into scene feature latent variables; The master encoder is used to encode the defensive driving trajectory sequence of an expert driver into latent trajectory feature variables; The decoder is used to concatenate scene feature latent variables and trajectory feature latent variables, and map them to Gaussian distribution parameters in the latent space, by sampling latent variables from the latent space. ,Will By fusing with latent variables of scene features, a planned trajectory for an active defense reference driver model is generated.

6. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 5, characterized in that, The attention module includes a temporal attention module and a target attention module. The temporal attention module is used to extract temporal feature vectors from scene state information. The target attention module is used to extract target interaction feature vectors from scene state information; The scene feature vector is obtained by concatenating the temporal feature vector with the target interaction feature vector.

7. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 1, characterized in that, S4 includes the following steps: S41. In typical defensive driving scenarios, the active defense reference driver model is tested multiple times to generate multiple reference trajectories, which are then mapped to a joint time-space coordinate system to define the reference spatiotemporal safety envelope region under a pre-set confidence level. ; S42. Under the same scenario conditions, test the autonomous driving system under test and record the corresponding spatiotemporal trajectory data.

8. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 7, characterized in that, S5 includes the following steps: S51. Calculate the safety score, efficiency score, and comfort score respectively; S52. Map the spatiotemporal trajectory of the tested autonomous driving system to a spatiotemporal coordinate system and determine the relationship between the tested trajectory and the reference safety envelope region. The inclusion relationship; S53. Calculate the comprehensive score of the defensive driving capability of the tested autonomous driving system.

9. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 8, characterized in that, The formula for calculating the security score is as follows: in, To proactively defend against the average risk magnitude faced by all reference trajectories of the reference driver model in the scenario, The magnitude of the risks faced by the tested autonomous driving system in the scenario. and All were normalized to the range [0,1]. As a correction factor, when The time characterization shows that the safety level of the tested autonomous driving system is on par with the active defense reference driver model; The formula for calculating the efficiency score is as follows: in, The average efficiency index of all reference trajectories of the active defense reference driver model is calculated. The results of efficiency index calculations for the tested autonomous driving system. This is the scaling factor for the limit value; The formula for calculating the comfort score is as follows: in, The average comfort index of all reference trajectories of the active defense reference driver model is calculated. The results of the comfort index calculation for the tested autonomous driving system. For comfort limits.

10. The method for evaluating the defensive driving capability of an autonomous driving system according to claim 8, characterized in that, The formula for calculating the comprehensive score of the defensive driving capability of the tested autonomous driving system is as follows: in, The overall score is for defensive driving ability. For safety rating, The spatiotemporal trajectory of the tested autonomous driving system. For efficiency scoring, For comfort rating, , and These are the weighting coefficients. and This is the scaling factor. The higher the value, the stronger the defensive driving ability.

Images