Method and System for Evaluating Driver Trust in L3-Level Autonomous Vehicles

Abstract

The present invention relates to a method for evaluating the driver trust level of an L3-level autonomous vehicle, which includes the following steps: Step 1, by introducing a subjective trust scale, obtain the driver trust level of the autonomous driving takeover experiment and conduct a quantitative analysis on the subjective trust level; Step 2, use the K-means clustering method to classify the subjective trust level; Step 3, extract the key frames of the driving takeover; Step 4, encode the extracted driver states; Step 5, establish a driver trust level evaluation method based on the Kalman filter algorithm. The present invention can provide a reference for the objective and real-time evaluation of the driver trust level in emergency takeover scenarios, which is beneficial to improving the safety and reliability of autonomous vehicles.

Description

Technical Field

[0001] The present invention belongs to the field of traffic safety, and particularly relates to a method and system for evaluating the driver trust level of an L3-level autonomous driving vehicle. Background Art

[0002] In recent years, autonomous driving has become a research hotspot in the traffic field. The Society of Automotive Engineers (SAE) has defined six levels of autonomous driving from 0 to 5, and L3-level represents conditional autonomous driving. When a vehicle in the autonomous driving state faces a dangerous emergency situation beyond its capabilities or is in a failure or malfunction state, the driver needs to perform emergency handling and take over the vehicle urgently. Under L3-level driving conditions, the driver often engages in secondary tasks unrelated to driving, and the awareness and response to the traffic environment will be reduced to varying degrees, resulting in a longer risk response time for the driver in the autonomous driving state than in manual driving, which is extremely likely to cause traffic conflicts or accidents.

[0003] In the existing methods for evaluating the driver trust level of autonomous driving vehicles, due to the characteristics of simplicity and easy implementation, the subjective scale method is widely adopted in research and extensively applied in this field. With the in-depth research, the limitations of the subjective trust scale method gradually emerge, and it is difficult to repeatedly obtain and update the trust evaluation results in real time over time. Therefore, there is an urgent need for a method and system for evaluating the driver trust level of an L3-level autonomous driving vehicle to provide a reference for the objective real-time evaluation of the driver trust level in the intelligent traffic environment. Summary of the Invention

[0004] The present invention proposes a method and system for evaluating the driver trust level of an L3-level autonomous driving vehicle to solve the above problems.

[0005] The present invention relates to a method for evaluating the driver trust level of an L3-level autonomous driving vehicle, including the following steps:

[0006] Step 1: By introducing a subjective trust scale, obtain the driver trust level of the autonomous driving takeover experiment and perform quantitative analysis on the subjective trust level;

[0007] Step 2: Use the K-means clustering method to classify the subjective trust level;

[0008] Step 3: Extract the key frames of the driving takeover;

[0009] Step 4: Encode the extracted driver state;

[0010] Step 5: Establish a method for evaluating the driver trust level based on the Kalman filter algorithm;

[0011] Further, in Step 1, the trust after the experiment is quantified based on the subjective trust scale, and the quantification method is the average value of the numbers of the subjective trust scale options for each subject.

[0012] Further, in Step 2, the K-means clustering method is used to divide the subjective trust degree into a high-trust group, a medium-trust group, and a low-trust group.

[0013] Further, in Step 3, the corresponding moments for selecting the takeover key frames are: 2.0 s before the takeover prompt, 1.0 s before the takeover prompt, 0.0 s at the takeover prompt, 1.0 s after the takeover prompt, and 2.0 s after the takeover prompt.

[0014] Further, in Step 4, the positions of the driver's hands and feet are classified and encoded as H1, H2, H3, F1, F2, and F3 respectively.

[0015] Further, in Step 5, the Kalman filter algorithm is divided into two modules: prediction update and measurement update, and the calculation formulas are as follows:

[0016]

[0017] In the formula, —— the system state at time t; F t —— the system transition matrix at time t; K t —— the Kalman gain at time t; Z t —— the actual observation value at time t; H t —— the observation matrix; P t —— the error matrix; R t —— the measurement noise covariance matrix; B t —— the input control matrix; u t —— the input control quantity; Q t —— the process noise covariance matrix;

[0018] Applying the classical continuous state estimation method based on the Kalman filter to evaluate the trust degree lies in that the continuous output metric of the estimator helps to construct a takeover performance prediction model considering the trust degree, laying a foundation for the design of the application controller and the decision algorithm; weakening the potential influence related to the randomness of the driver's behavior through the Kalman filter with repeated measurement efficiency to achieve real-time and accurate evaluation of the driver's trust degree;

[0019] To effectively carry out the establishment of the method, the warning type corresponding to the takeover event t is converted using boolean values, and the conversion method is as follows:

[0020] ① Correct warning

[0021]

[0022] ②Error warning

[0023]

[0024] ③No warning

[0025]

[0026] The implementation of the Kalman filter requires the definition of observation variables that can be measured and processed in real time. These observation variables must be related to the variables to be evaluated (i.e., the trust level). Therefore, the frequency hi when the driver's hand position is H1 and the frequency fi when the foot position is F1 are defined as the observation variables in the Kalman filter. At the same time, combining the subjective trust level and the warning type boolean value, an LTI system state space model is used to represent the driver's dynamic trust level during the experiment. The formula is as follows:

[0027]

[0028] In the formula, T o (t + 1) —— The objective trust level of the subject after the (t + 1)-th takeover;

[0029] T o (t) —— The objective trust level of the subject after the t-th takeover;

[0030] J t+1 ,L t+1 ,M t+1 —— The boolean value of the warning type for the (t + 1)-th takeover event;

[0031] h t+1 —— The frequency of the driver's hand position at level H1 in the (t + 1)-th takeover event;

[0032] f t+1 —— The frequency of the driver's foot position at level F1 in the (t + 1)-th takeover event;

[0033] A, B, C —— Undetermined coefficient matrices, where A = [a 11 , B = [b 11 , b 12 , b 13 , C = [c 11 , c 21 T ;

[0034] u(t) —— The noise in the prediction process,

[0035] w(t) —— The observation noise, w(t) ~ Ν(0, ∑w);

[0036] ​The Kalman filter algorithm is used to evaluate the trust based on the state space equation. The algorithm flow is as follows:

[0037] (1) Initialize the trust and covariance

[0038]

[0039] In the formula, ——The initial objective prior trust of the driver evaluated by the algorithm, which can be evaluated by the trust scale in practical applications;

[0040] C——Undetermined coefficient matrix, C=[c 11 ,c 21 ] T ;

[0041] h1——the frequency of the driver’s hand position being at level H1 in the first takeover event;

[0042] f1 – the frequency of the driver’s foot position being at level F1 during the first takeover event;

[0043] —— The covariance of

[0044] (2) Calculation of Kalman gain

[0045]

[0046] Where, K is Kalman gain;

[0047] —— The covariance of

[0048] ∑w——standard deviation of observation noise;

[0049] (3) Driver status update

[0050]

[0051] In the formula, ——The frequency of the driver’s hand position being at level H1 as assessed by the algorithm for the t+1 takeover event;

[0052] ——The frequency of the driver’s foot position being F1 as assessed by the algorithm for the t+1 takeover event;

[0053] ——the driver’s prior objective trust evaluated by the algorithm after t takeover events;

[0054]

[0055] In the formula, ht+1 —— The frequency of the driver's hand position at the (t + 1)-th takeover event being at level H1;

[0056] f t+1 —— The frequency of the driver's foot position at the (t + 1)-th takeover event being at level F1;

[0057] (4) Driver trust update

[0058]

[0059] In the formula, T o (t)—— The posterior objective trust of the driver evaluated by the algorithm after the t-th takeover;

[0060] ∑T o (t)—— The covariance of T o (t);

[0061]

[0062] In the formula, —— The prior objective trust of the driver evaluated by the algorithm after the (t + 1)-th takeover event;

[0063] —— The covariance of;

[0064] σ u —— The standard deviation of the predicted process noise.

[0065] The present invention also relates to a driver trust evaluation system for an L3-level autonomous driving vehicle applying the above method. The system includes a subjective trust evaluation module, a driver state feature extraction device, and a driver trust evaluation module.

[0066] The subjective trust evaluation module is a subjective trust scale filled in by the driver before and after driving; the reliability level of the subjective trust scale is obtained by calculating the Cronbach’s alpha coefficient, and auxiliary analysis is carried out to obtain the subjective trust of the driver of the L3-level autonomous driving vehicle.

[0067] The driver state feature extraction device includes: a Logitech G29 steering wheel, a throttle and brake pedal kit, a stereo, a takeover simulation environment display, and a video acquisition device; the Logitech G29 steering wheel, the throttle and brake pedal kit, the stereo, and the takeover simulation environment display are used to implement the driver's simulated driving operation, and the video acquisition device is used to collect the hand and foot features of the driver.

[0068] The driver trust evaluation module is a Kalman filter for evaluating driver trust; after the state space equation based on the driver's hand and foot features is input into the Kalman filter, the objective trust of the L3-level autonomous driving vehicle driver based on the driver's state features is obtained.

[0069] Beneficial effects

[0070] Based on the takeover simulation experiment under L3-level autonomous driving conditions and combined with the subjective trust evaluation results, the present invention establishes a real-time trust evaluation method based on the driver's hand and foot states and quantifies and grades the driver's trust. The research results provide a reference for the objective and real-time evaluation of driver trust in emergency takeover scenarios, which is beneficial to improving the safety and reliability of autonomous driving vehicles. Description of the drawings

[0071] Figure 1 It is a schematic structural diagram of the driver trust evaluation system for the L3-level autonomous driving vehicle of the present invention. Detailed implementation manners

[0072] The following is a specific description in combination with Figure 1 This specific implementation manner will be specifically described.

[0073] The method and system for evaluating the driver trust of the L3-level autonomous driving vehicle of the present invention include the following steps:

[0074] Step 1: By introducing a subjective trust scale, obtain the driver trust in the autonomous driving takeover experiment and conduct a quantitative analysis of the subjective trust.

[0075] To obtain the driver trust after the real vehicle driving experiment, a and other developed subjective trust questionnaires were used to evaluate the driver's subjective trust. The structure of this questionnaire is divided into 5 sub-scales (reliability, familiarity, trust, comprehensibility, and developer intention) and 19 questions. The scale uses a 5-level Likert scale, which can achieve a detailed evaluation of the driver's subjective trust level from multiple dimensions.

[0076] Quantify the trust after the experiment based on the subjective trust scale. The quantification method is the numerical average of the subjective trust scale options of each subject (option transposition should be performed for some questions).

[0077] During the experiment, there were cases of false warnings and non-warnings in the system to regulate the subjects' trust in the system. To collect the changes in the subjects' trust during the experiment, after each takeover operation, the subjects would fill out a simple multiple-choice question with options {significantly decreased, slightly decreased, unchanged, slightly increased, significantly increased} to collect the dynamic changes in the subjects' trust during the experiment. To unify the subjects' post-experiment trust and the dynamic trust during the experiment, the subjective trust of the subjects was defined as:

[0078]

[0079] where T s (t) —— the subjective trust of the subject after the t-th takeover;

[0080] T sp —— the subjective trust of the subject after the experiment;

[0081] ΔT s (t + 1) —— the change level of the subjective trust of the subject after the (t + 1)-th takeover, and its values are {-2, -1, 0, 1, 2}, corresponding to {significantly decreased, slightly decreased, unchanged, slightly increased, significantly increased} respectively.

[0082] Step 2: Use the K-means clustering method to classify the subjective trust.

[0083] Use the K-means clustering method to divide the subjective trust of the driver in the autonomous driving system into different levels. The K-means clustering algorithm is a commonly used unsupervised clustering method. It can divide n trust evaluation results (T1, T2,..., Tn) into k clusters with the criterion of minimizing the within-cluster variance. Each trust evaluation result belongs to the nearest cluster center (the average value of each cluster). The steps of the K-means algorithm are as follows: (1) Determine the number of clusters; (2) Select an initial cluster center for each cluster; (3) Assign the data in the dataset to the nearest cluster according to the minimum distance principle; (4) Calculate the trust mean of each cluster and update the cluster center; (5) Repeat the assignment of categories and update the cluster center until the categories to which each sample belongs no longer change; (6) Output the final cluster center and the k-cluster division.

[0084] Based on the above steps and the principle of the K-means algorithm, the trust quantification results in Step 2 are classified. First, the value of the algorithm k, that is, the number of classifications, should be determined. The commonly used methods for selecting the value of k are the elbow method and the silhouette coefficient method. The principle of the elbow method is to calculate the sum of squared errors of all sample points in the dataset to their cluster centers, and select the value (the inflection point of the curve) when the sum of squared errors suddenly becomes smaller as the value of k of the algorithm. The silhouette coefficient method selects the value of k by calculating the silhouette coefficient. The larger the average value of the silhouette coefficients of all sample points, the better the clustering effect. The calculation formula for the silhouette coefficient of each sample point is as follows:

[0085]

[0086] In the formula, s i —— The silhouette coefficient of sample point i in a certain cluster;

[0087] a i —— The average distance from sample point i to other sample points belonging to the same cluster. The smaller a i is, the greater the possibility that sample point i belongs to this category;

[0088] b i —— The minimum value of the average distance from sample point i to all samples in other clusters.

[0089] The inflection point coordinates of the elbow method curve are 3, indicating that the optimal number of clusters for the study is 3. The maximum value point of the silhouette coefficient in the silhouette coefficient method is 7, indicating that the optimal number of clusters is 7. Considering that too many or too few clusters are not conducive to analyzing the impact of trust degree changes on takeover performance indicators and constructing a prediction model for takeover performance. Therefore, in this study, the optimal number of clusters is taken as 3, and the K-means algorithm is used to divide the driver's subjective trust degree into three different trust groups: high trust degree, medium trust degree, and low trust degree. The number of samples is 150 (30 * 5).

[0090] From the clustering calculation results, it can be seen that among the 150 subjective trust degree evaluation samples, 56 samples are clustered into the low trust group, and the trust degree of its cluster center is 2.80; 58 samples are clustered into the medium trust group, and its cluster center is 3.99; 36 samples are clustered into the high trust group, and its cluster center is 5.26.

[0091] To further analyze the differences between different trust groups, first, a normality test is performed on the trust levels of the three trust groups. The results show that at the significance level of 0.05, the samples do not significantly come from a normally distributed population. Therefore, the Firedman non-parametric test method is used to test the differences between groups. This method does not require the assumption of normal distribution. The test results are shown in Table 1.

[0092] Table 1 Statistical test results between different trust groups

[0093]

[0094] As can be seen from Table 1, the test results show that there are significant differences among the three trust groups. There are significant differences (p<0.01) between the low-trust group (2.80±0.06) and the medium-trust group (3.99±0.06), and between the low-trust group and the high-trust group (5.26±0.08). There is also a significant difference (p<0.01) between the medium-trust group (3.99±0.06) and the high-trust group (5.26±0.08).

[0095] Step 3: Extract the key frames of driving takeover.

[0096] The corresponding moments for selecting the takeover key frames are: 2.0 s before the takeover prompt, 1.0 s before the takeover prompt, 0.0 s at the takeover prompt, 1.0 s after the takeover prompt, and 2.0 s after the takeover prompt.

[0097] Step 4: Encode the extracted driver states.

[0098] Encode the hand positions in the five takeover key frames for each takeover event. The position of each hand can be divided into a relaxed state, a hovering state, and a controlling state. The relaxed state means that the subject's hand is not on the steering wheel and is not ready to control and input to the vehicle. The hovering state means that the subject's hand is not on the steering wheel but is ready to control and input to the vehicle. The controlling state means that the subject's hand is on the steering wheel.

[0099] According to the degree of driver involvement during driving, the combinations of the positions of both hands are divided into the following three different categories: H1 (both hands are in the relaxed state), H2 (one hand is in the hovering state and the other hand is in the hovering or relaxed state), and H3 (at least one hand is in the controlling state). Classify and encode the position of the driver's feet as: F1 (the right foot is in the relaxed state), F2 (the right foot is in the hovering state), and F3 (the right foot is in the controlling state).

[0100] Step 5: Establish a driver trust evaluation method based on the Kalman filtering algorithm.

[0101] Kalman Filtering is an algorithm for estimating the optimal state of a system. It uses a linear system state equation to fuse the input observed data and estimated data, filters the noise and interference in the system, and manages the error in a closed loop, enabling the algorithm to maintain a relatively small error in long-term sequence operations and finally output the optimal system state.

[0102] The Kalman filtering algorithm is mainly divided into two modules: prediction update and measurement update. The calculation formulas are as follows:

[0103]

[0104] In the formula, —— the system state at time t; F t —— the system transition matrix at time t; K t —— the Kalman gain at time t; Z t —— the actual observed value at time t; H t —— the observation matrix; P t —— the error matrix; R t —— the measurement noise covariance matrix; B t —— the input control matrix; u t —— the input control quantity; Q t —— the process noise covariance matrix;

[0105] Applying the classical continuous state estimation method based on the Kalman filter to evaluate the trust level lies in that the continuous output metric of the estimator helps to construct a takeover performance prediction model considering the trust level and at the same time lays a foundation for the design of controllers and decision algorithms in future applications. Secondly, the Kalman filter with repeated measurement effectiveness can weaken the potential impact related to the randomness of driver behavior and achieve real-time and accurate evaluation of driver trust level.

[0106] To effectively carry out the establishment of the method, first, the warning types corresponding to the takeover event t are converted using boolean values, and the conversion method is as follows:

[0107] ① Correct warning

[0108]

[0109] ② False warning

[0110]

[0111] ③ No warning

[0112]

[0113] The implementation of the Kalman filter requires defining observation variables that can be measured and processed in real time. These observation variables must be related to the variables to be evaluated (i.e., the trust level). Therefore, the frequency h when the driver's hand position is H1 i and the frequency f when the foot position is F1 i are defined as the observation variables in the Kalman filter. At the same time, combining the subjective trust level and the boolean value of the warning type, the LTI system state space model is used to represent the driver's dynamic trust level during the experiment. The formula is:

[0114]

[0115] In the formula, To (t + 1)——The objective trust level of the subject after the (t + 1)-th takeover;

[0116] T o (t)——The objective trust level of the subject after the t-th takeover;

[0117] J t+1 , L t+1 , M t+1 ——The Boolean value of the warning type of the (t + 1)-th takeover event;

[0118] h t+1 ——The frequency that the driver's hand position is at H1 level in the (t + 1)-th takeover event;

[0119] f t+1 ——The frequency that the driver's foot position is at F1 level in the (t + 1)-th takeover event;

[0120] A, B, C——Coefficient matrices to be determined, where A = [a 11 , B = [b 11 , b 12 , b 13 , C = [c 11 , c 21 ; T ;

[0121] u(t)——The noise of the prediction process,

[0122] w(t)——The observation noise, w(t) ~ Ν(0, ∑ w ).

[0123] Subsequently, the trust level is evaluated using the Kalman filter algorithm based on the state space equation, and the algorithm flow is as follows:

[0124] (1) Initialize the trust level and covariance

[0125]

[0126] In the formula, ——The initial objective prior trust level of the driver evaluated by the algorithm, which can be evaluated using a trust scale in practical applications;

[0127] C——Coefficient matrix to be determined, C = [c 11 , c 21 ; T ;

[0128] h1——The frequency that the driver's hand position is at H1 level in the 1st takeover event;

[0129] f1——The frequency that the driver's foot position is at F1 level in the 1st takeover event;

[0130] —— Covariance of

[0131] (2) Calculate the Kalman gain

[0132]

[0133] In the formula, K——Kalman gain;

[0134] —— Covariance of

[0135] ∑ w ——Standard deviation of observation noise.

[0136] (3) Driver state update

[0137]

[0138] In the formula, ——Frequency that the driver's hand position is at H1 level evaluated by the takeover event algorithm at the (t + 1)th time;

[0139] ——Frequency that the driver's foot position is at F1 level evaluated by the takeover event algorithm at the (t + 1)th time;

[0140] ——Driver's prior objective trust level evaluated by the algorithm after the tth takeover event.

[0141]

[0142] In the formula, h t+1 ——Frequency that the driver's hand position is at H1 level at the (t + 1)th takeover event;

[0143] f t+1 ——Frequency that the driver's foot position is at F1 level at the (t + 1)th takeover event.

[0144] (4) Driver trust level update

[0145]

[0146] In the formula, T o (t)——Driver's posterior objective trust level evaluated by the algorithm after the tth takeover.

[0147] ∑T o (t)——Covariance of To(t).

[0148]

[0149] Wherein, —— The prior objective trust level of the driver evaluated by the algorithm after the (t + 1)-th takeover event;

[0150] —— The covariance of

[0151] σ u —— The standard deviation of the predicted process noise.

[0152] The present invention designs a driver trust level evaluation system for L3-level autonomous driving vehicles by applying the above method. The system includes a subjective trust level evaluation module, a driver state feature extraction device, and a driver trust level evaluation module.

[0153] The subjective trust level evaluation module is a subjective trust scale filled in by the driver before and after driving; the reliability level of the subjective trust scale is obtained through the calculated Cronbach's alpha coefficient for auxiliary analysis to obtain the subjective trust level of the driver of the L3-level autonomous driving vehicle.

[0154] The driver state feature extraction device includes: a Logitech G29 steering wheel, a throttle and brake pedal kit, a stereo, a takeover simulation environment display, and a video acquisition device; the Logitech G29 steering wheel, the throttle and brake pedal kit, the stereo, and the takeover simulation environment display are used to implement the driver's simulated driving operation, and the video acquisition device is used to acquire the hand and foot features of the driver.

[0155] The driver trust level evaluation module is a Kalman filter for evaluating the driver's trust level; after the state space equation based on the driver's hand and foot features is input into the Kalman filter, the objective trust level of the driver of the L3-level autonomous driving vehicle based on the driver's state features is obtained.

[0156] Embodiment

[0157] In this embodiment, the SCANeR studio software is used as the driving simulation platform to construct a simulation scenario, control the parameters of the host vehicle and other vehicles, and realize the switching between autonomous driving and manual driving. During the experiment, the automatic transmission mode is adopted when the host vehicle needs to take over, and the driver only needs to operate the throttle, brake pedal, and steering wheel during manual driving. During the simulated driving process, the software can record vehicle data in real time, including vehicle speed, acceleration, steering wheel angle, brake pedal change curve, throttle opening degree, and other data.

[0158] The experimental scenario is designed as a 12-km long straight section of highway. According to the "Technical Standard for Highway Engineering" (JTGB01-2014), the design parameters of this section are as follows: the speed limit of the section is 120 km / h, the vehicle speed during autonomous driving is set at 110 km / h, the lane is designed as an eight-lane two-way lane, the width of each lane is 3.75 m, the traffic flow is set as a stable flow (15 pcu / km·ln), and the weather is set as sunny. A total of four takeover scenarios are set in the formal experiment. The warning types of the scenarios are correct warning, false warning, correct warning, and no warning respectively, and the appearance order of the warning types is different from that of the pre-experiment. The distance between the trigger points of different takeover scenarios is 2.5 km.

[0159] To obtain sufficient experimental data for the study of driver trust evaluation, 30 drivers were recruited as subjects. The average age of the subjects was 21.2 years, the average driving experience was 0.42 years, and the average driving mileage was 240.67 km. By comparing and analyzing the accuracy rate, F1 value, Recall value and other indicators of machine learning algorithms such as XGBoost and random forest, it was found that the XGBoost algorithm has better prediction accuracy, and its various evaluation indicators are greater than those of the other algorithms, thus verifying the effectiveness of the XGBoost algorithm.

[0160] In this embodiment, the above content is only the preferred embodiment of the present invention and is not used to limit the implementation of the present invention. Those of ordinary skill in the art can easily make changes or modifications according to the main concept and spirit of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope required by the claims.

Claims

1. A method for evaluating the driver trust level of an L3 - level autonomous driving vehicle, characterized in that, It includes the following steps: Step 1: By introducing a subjective trust scale, obtain the driver trust level of the autonomous driving takeover experiment, and conduct a quantitative analysis of the subjective trust level; quantify the trust after the experiment based on the subjective trust scale, and the quantification method is the numerical average of the subjective trust scale options for each subject; Step 2: Use the K-means clustering method to classify the subjective trust level; Step 3: Extract the key frames of the driving takeover, and the corresponding moments selected for the takeover key frames are: 2.0 s before the takeover prompt, 1.0 s before the takeover prompt, 0.0 s at the takeover prompt, 1.0 s after the takeover prompt, and 2.0 s after the takeover prompt; Step 4: Encode the extracted driver states, and classify and encode the positions of the driver's hands and feet as H1, H2, H3, F1, F2, F3 respectively; Step 5: Establish a driver trust level evaluation method based on the Kalman filter algorithm.

2. The method for evaluating the driver trust level of an L3 - level autonomous driving vehicle according to claim 1, characterized in that, In Step 2, the K-means clustering method is used to divide the subjective trust level into a high trust group, a medium trust group, and a low trust group.

3. The method for evaluating the driver trust level of an L3 - level autonomous driving vehicle according to claim 1, characterized in that, In Step 5, the Kalman filter algorithm is divided into two modules: prediction update and measurement update, and the calculation formula is as follows: Wherein, —— The system state at time t; F t —— The system transition matrix at time t; K t —— The Kalman gain at time t; Z t —— The actual observed value at time t; H t —— The observation matrix; P t —— The error matrix; R t —— The measurement noise covariance matrix; B t —— The input control matrix; u t —— The input control quantity; Q t —— The process noise covariance matrix; Use a boolean value to convert the warning type corresponding to the takeover event t, and the conversion method is as follows: ① Correct warning ② False warning ③ No warning The implementation of the Kalman filter requires the definition of observation variables that can be measured and processed in real time. These observation variables must be related to the variables to be evaluated. Therefore, the frequency h when the driver's hand position is H1 i and the frequency f when the foot position is F1 i are defined as the observation variables in the Kalman filter. At the same time, combining the subjective trust level and the warning type boolean value, an LTI system state space model is used to represent the driver's dynamic trust level during the experiment. The formula is: where, T o (t + 1)——the objective trust level of the subject after the (t + 1)-th takeover; T o (t) —— The objective trust level of the subject after the t-th takeover; J t+1 ,L t+1 ,M t+1 —— Boolean value of the early warning type of the (t + 1)-th takeover event; h t+1 —— The frequency that the driver's hand position is at H1 level in the (t + 1)-th takeover event; f t+1 —— the frequency of the driver's foot position being at F1 level in the (t + 1)-th takeover event; A, B, C——Undetermined coefficient matrices, where A = [a 11 , B = [b 11 , b 12 , b 13 , C = [c 11 , c 21 T ;​ u(t) —— Noise in the prediction process, w(t) - Observation noise, w(t) ~ Ν(0,∑w); Based on the state space equation, use the Kalman filter algorithm to evaluate the trust level, and the algorithm flow is as follows: (1) Initialize the trust level and covariance In the formula, —— The initial objective prior trust degree of the driver evaluated by the algorithm, which can be evaluated by a trust scale in practical applications; C——Coefficient matrix to be determined, C = [c 11 , c 21 T ;​ h1 - The frequency of the driver's hand position being at level H1 in the first takeover event; f1 - The frequency of the driver's foot position being at level F1 in the first takeover event; —— Covariance of; (2) Calculate the Kalman gain In the formula, K - Kalman gain; —— Covariance of; ∑w - Standard deviation of the observation noise; (3) Update the driver state In the formula, —— the frequency that the driver's hand position is at H1 level in the algorithm evaluation of the takeover event at the (t + 1)th time; ——The frequency that the driver's foot position is at F1 level in the algorithm evaluation of the takeover event at t+1 times; ——The prior objective trust of the driver evaluated by the algorithm after the t-th takeover event; where h t+1 —— the frequency that the driver's hand position at the (t + 1)-th takeover event is at H1 level; f t+1 ——The frequency that the foot position of the driver in the (t + 1)-th takeover event is at F1 level; (4) Update the driver trust level where, T o (t)——the posterior objective trust degree of the driver evaluated by the algorithm after the t-th takeover; ∑T o (t)——T o (t) covariance; wherein, —— the prior objective trust degree of the driver evaluated by the algorithm after the (t + 1)-th takeover event; —— the covariance of; σ u —— Standard deviation of the predicted process noise.

4. A system for implementing the method for evaluating the driver trust level of an L3 - level autonomous driving vehicle according to any one of claims 1 to 3, characterized in that, The system includes a subjective trust level evaluation module, a driver state feature extraction device, and a driver trust level evaluation module.

5. The system for the method for evaluating the driver trust level of an L3 - level autonomous driving vehicle according to claim 4, characterized in that, The subjective trust level evaluation module is the subjective trust scale filled in by the driver before and after driving; Obtain the reliability level of the subjective trust scale through the calculated Cronbach’s alpha coefficient, conduct auxiliary analysis, and obtain the subjective trust level of the L3-level autonomous driving vehicle driver.

6. The system for the method for evaluating the driver trust level of an L3 - level autonomous driving vehicle according to claim 4, characterized in that, The driver state feature extraction device includes: Logitech G29 steering wheel, throttle and brake pedal kit, audio, takeover simulation environment display, video acquisition device; the Logitech G29 steering wheel, throttle and brake pedal kit, audio, and takeover simulation environment display are used to implement the driver's simulated driving operation, and the video acquisition device is used to collect the hand and foot features of the driver.

7. The system for the method for evaluating the driver trust level of an L3 - level autonomous driving vehicle according to claim 4, characterized in that, The driver trust level evaluation module is a Kalman filter used to evaluate the driver trust level; after the state space equation based on the driver's hand and foot features is input into the Kalman filter, the objective trust level of the L3-level autonomous driving vehicle driver based on the driver state features is obtained.

Images