A method for recognizing driver trust state and behavior verification in pilot assisted driving

Abstract

The application discloses a kind of pilot trust state in piloting auxiliary driving identification and behavior verification method;Including: collecting the physiological signal data and phased subjective evaluation data of multiple drivers under piloting auxiliary driving task;Characteristic extraction and normalization processing are carried out to physiological signal, and heart rate variability related feature and galvanic skin response related feature are obtained;Two kinds of features are respectively reduced dimension and fused using grouping principal component analysis method, and heart rate variability comprehensive score and galvanic skin response comprehensive score are calculated;Subjective trust score and two physiological comprehensive scores are combined, and the comprehensive trust degree score of each driving task is weighted calculated;Driving task is divided into high and low trust degree state based on clustering algorithm;Extract corresponding vehicle operation data, analyze the behavior difference in safety, comfort, efficiency index under different states, and verify the rationality of state division by statistical method.The objective quantification of pilot trust state and the verifiability of behavior level are realized.

Description

Technical Field

[0001] This invention relates to the field of navigation-assisted driving, and more particularly to a method for identifying and verifying the driver's trust state and behavior in navigation-assisted driving. Background Technology

[0002] With the rapid development of smart cockpit and autonomous driving technologies, automakers such as Tesla, Huawei, and XPeng have widely deployed L2 / L2+ level driver assistance functions. During use, drivers not only focus on driving safety but also increasingly value ride comfort and system reliability. Their trust in the driver assistance system has become a key factor influencing the frequency of function activation and user acceptance. Currently, driver trust assessment relies heavily on subjective evaluation methods, such as phased questionnaires and post-trip interviews. While these methods directly reflect user psychological perception, they suffer from limitations such as strong subjectivity, significant individual differences, and difficulty in quantification and real-time monitoring. To improve the objectivity and repeatability of assessments, recent studies have attempted to introduce physiological signals as supplementary indicators, such as heart rate variability and skin conductance, to reflect the driver's cognitive load and emotional state. However, single-modal data cannot comprehensively characterize the dynamic changes in trust and lacks correlation verification with actual driving behavior. Existing methods have not yet achieved effective integration of physiological signals and subjective feedback, nor have they verified the quantified trust status at the behavioral level. Therefore, there is an urgent need for a multimodal trust quantification method that integrates subjective evaluation and physiological signals. This method can be used to classify high / low trust states through clustering and verify the rationality of the classified states based on the behavioral differences of vehicle operation data in normal and special driving scenarios. This will enable the construction of an interpretable and verifiable driver trust state identification and evaluation system. Summary of the Invention

[0003] To effectively integrate physiological signals and subjective feedback, and to verify the quantified trust level at the behavioral level, this invention proposes a method for identifying and verifying the driver's trust level in assisted driving, comprising the following steps:

[0004] S1: Collect physiological signal data and subjective evaluation data of multiple drivers under corresponding driving tasks. The subjective evaluation data is obtained through a phased questionnaire. Calculate the driver's subjective trust score for the corresponding driving task based on the subjective evaluation data. The driving task belongs to the navigation-assisted driving process.

[0005] S2: Feature extraction and normalization are performed on physiological signal data to obtain heart rate variability-related features and skin conductance response-related features. Grouped principal component analysis is used to reduce the dimensionality of the heart rate variability-related features and skin conductance response-related features and calculate the comprehensive heart rate variability score and comprehensive skin conductance response score corresponding to the heart rate variability-related features and skin conductance response-related features, respectively.

[0006] S3: The driver's subjective trust score, the comprehensive score of heart rate variability, and the comprehensive score of skin conductance response are weighted and calculated to obtain the comprehensive trust score for the corresponding driving task.

[0007] S4: Based on the comprehensive trust score of each driving task, a clustering algorithm is used to divide each driving task into a high trust state or a low trust state.

[0008] S5: Extract the vehicle operation data corresponding to the driving task, analyze the differences in driving behavior under different trust levels based on the vehicle operation data, and use statistical methods to determine whether the differences are significant in order to verify the consistency between the classified trust levels and the actual driving behavior.

[0009] Furthermore, the heart rate variability-related features include: the average heart rate during a single driving task, the mean RR interval, and the ratio of total power in the low-frequency band to total power in the high-frequency band in the frequency domain analysis of heart rate variability;

[0010] The skin conductance response-related features include: the average amplitude, average rise time, average recovery time, and average frequency of SCR events per minute for all valid SCR events during a single driving task.

[0011] The effective SCR event refers to a skin conductance response event in which the change in conductance between the peak and baseline of the SCR event exceeds 0.01 μs and has a typical single-peak morphology, with a rise time between 0.5 and 5 seconds and a recovery time longer than the rise time.

[0012] Furthermore, the calculation method for the heart rate variability composite score or skin conductance response composite score includes:

[0013] Modeling phase: Principal component analysis is performed on the target-related feature data of multiple driving tasks to obtain multiple eigenvalues, eigenvectors corresponding to each eigenvalue, the number of retained eigenvectors (i.e., the number of principal components), and the principal component direction matrix formed by the combination of eigenvectors corresponding to the retained principal components; the principal component refers to the direction represented by the eigenvector.

[0014] Application phase: Using the principal component direction matrix obtained in the modeling phase, linear transformation is performed on the feature vectors corresponding to each principal component in the current driving task to obtain the score corresponding to each principal component; the scores corresponding to each principal component are preprocessed.

[0015] The total variance information corresponding to the target-related feature data in the current driving task is calculated based on the retained principal components, and the weight corresponding to the target-related feature data is calculated based on the total variance information.

[0016] The comprehensive score corresponding to the target-related feature data is calculated based on the scores of each principal component after preprocessing and the weights of the target-related feature data.

[0017] The target-related feature data are heart rate variability-related features or skin conductance response-related features; the comprehensive score corresponding to the target-related feature data is the comprehensive score of heart rate variability and the comprehensive score of skin conductance response.

[0018] Furthermore, the modeling stage specifically includes:

[0019] Based on the relevant feature data of multiple driving tasks, a multidimensional feature matrix is ​​formed; the covariance matrix of the multidimensional feature matrix is ​​calculated, and its eigenvalues ​​are decomposed to obtain multiple eigenvalues ​​and their corresponding eigenvectors.

[0020] For each eigenvalue, calculate the ratio of that eigenvalue to the sum of all eigenvalues ​​to obtain the variance contribution rate corresponding to that eigenvalue;

[0021] Arrange the eigenvalues ​​in descending order and accumulate the variance contribution rates corresponding to the eigenvalues ​​in turn. When the accumulated value reaches a preset threshold, the number of eigenvalues ​​or eigenvectors contained therein is the number of principal components retained.

[0022] The principal component direction matrix is ​​formed by combining the eigenvectors corresponding to the retained principal components.

[0023] Furthermore, the application phase specifically includes:

[0024] The principal component direction matrix obtained in the modeling stage is used to perform a linear transformation on the feature vectors corresponding to each principal component in the current driving task to obtain the score corresponding to each principal component.

[0025] Calculate the Pearson correlation coefficient between the scores of each principal component and the driver's subjective trust score corresponding to the same driving task; if the correlation coefficient is negative, then perform numerical inversion on the principal component score; so that the higher the adjusted scores of all principal components, the higher the corresponding driver trust level.

[0026] Furthermore, the principal components are sorted from largest to smallest according to their corresponding eigenvalues. The first principal component corresponds to the largest eigenvalue, representing the main direction of data change. The eigenvectors corresponding to each subsequent principal component are orthogonal to the eigenvectors corresponding to all previous principal components, and their eigenvalues ​​are the largest among the remaining directions.

[0027] Furthermore, based on the comprehensive trust score for each driving task, a clustering algorithm is used to classify each driving task into a high-trust state or a low-trust state; specifically:

[0028] Randomly select k ratings from the overall trust scores of all driving tasks as the initial cluster centers;

[0029] Each data point is assigned to the nearest cluster center to form multiple clusters. The centroid of each cluster is recalculated, and the process is iterated until the centroid stabilizes or the maximum number of iterations is reached. Each data point corresponds one-to-one with the comprehensive trust score of the driving task.

[0030] The optimal number of clusters, k=2, was determined using the silhouette coefficient method and the elbow rule.

[0031] Based on the final clustering results, the cluster with the higher centroid value is defined as a high-trust state, and the cluster with the lower centroid value is defined as a low-trust state.

[0032] Each driving task is classified into a corresponding trust level state based on the cluster it belongs to.

[0033] Furthermore, the vehicle operation data includes safety indicators, comfort indicators, and efficiency indicators collected during the following time periods:

[0034] Vehicle operation data corresponding to the entire process of a single driving mission;

[0035] Vehicle operation data corresponding to each preset driving scenario in a single driving task. The preset driving scenarios include regular driving scenarios and special driving scenarios. Among them, regular driving scenarios include left-turn sections and right-turn sections; special driving scenarios include construction area sections and main-auxiliary road switching sections.

[0036] Furthermore, the security indicators include:

[0037] Minimum collision time, which represents the minimum value among the estimated collision time between the vehicle and the target ahead, is obtained by continuously monitoring the relative distance and relative speed and calculating the collision time at each moment and taking the minimum value.

[0038] Weighted arrival time is calculated by assigning weights to the estimated collision time at each moment based on risk factors, and then calculating it through weighted averaging or weighted comprehensive; it is used to reflect the overall potential collision risk level of the driving task; the risk factors include relative distance and relative speed.

[0039] Minimum relative distance, which represents the minimum relative distance between the vehicle and the target ahead;

[0040] The comfort indicators include: maximum longitudinal acceleration, minimum longitudinal acceleration, absolute mean lateral acceleration, standard deviation of heading angle, and the maximum absolute value of the rate of change of longitudinal acceleration, i.e., peak jerk.

[0041] The efficiency metrics include: average driving speed and speed standard deviation.

[0042] Furthermore, step S5 specifically includes:

[0043] Driving tasks in high-trust and low-trust states are categorized into high-trust and low-trust groups respectively.

[0044] For both regular and special driving scenarios, the values ​​of safety, comfort, and efficiency indicators corresponding to each driving task in the two groups are extracted respectively.

[0045] For each indicator, a comparison was made between the high-confidence group and the low-confidence group: the means of the two groups were calculated, and statistical hypothesis testing methods were used to determine whether the difference was significant.

[0046] If, under normal driving scenarios, the mean standard deviation of longitudinal acceleration in the high-trust group is significantly smaller than that in the low-trust group, and the mean minimum collision time is significantly larger than that in the low-trust group; and under special driving scenarios, the mean minimum relative distance in the high-trust group is significantly smaller than that in the low-trust group, then the classified trust status is determined to be consistent with the actual driving behavior.

[0047] Compared with the prior art, the present invention has at least the following beneficial effects:

[0048] (1) This invention achieves a multimodal comprehensive assessment of trust level during navigation-assisted driving by integrating driver subjective evaluation data and physiological signal data. Compared with the traditional subjective evaluation method that relies solely on questionnaires, this invention combines the driver's real psychological feedback and objective physiological response, effectively reducing the influence of individual subjective bias and improving the comprehensiveness and reliability of trust level assessment.

[0049] (2) This invention performs feature extraction and normalization on physiological signal data, and introduces grouped principal component analysis to perform dimensionality reduction and fusion on heart rate variability-related features and skin conductance response-related features, respectively, to generate a comprehensive score for heart rate variability and a comprehensive score for skin conductance response. This method avoids the interference of differences in the dimensions of different physiological indicators on the results, retains the main information features, improves the stability and interpretability of feature expression, and provides high-quality input for subsequent trust quantification.

[0050] (3) In the application stage of principal component analysis, this invention calculates the Pearson correlation coefficient between principal component scores and subjective trust scores, and reverses the values ​​of negatively correlated scores to ensure that the numerical direction of all comprehensive scores is consistent with the level of trust. This processing method solves the problem of inconsistent direction between physiological indicators and trust, making the fusion results more consistent with cognitive logic and enhancing the interpretability of the model output.

[0051] (4) Based on the comprehensive trust score of each driving task, this invention uses a clustering algorithm to automatically classify driving tasks into high-trust or low-trust states, and determines the optimal number of clusters k=2 using the silhouette coefficient method and the elbow rule, thus achieving objective and unsupervised classification of trust states. This method does not require a pre-set threshold, avoids the subjectivity brought about by manual classification, and improves the scientificity and consistency of state recognition.

[0052] (5) This invention further extracts vehicle operation data corresponding to driving tasks, analyzes the differences in driving behavior under different trust levels, and judges the significance of the differences through statistical hypothesis testing, thereby realizing the external behavioral verification of the classified trust levels. This verification mechanism establishes the correlation between psychological state and actual driving behavior, enhancing the credibility and practicality of the trust assessment results.

[0053] (6) In the analysis of vehicle operation data, this invention distinguishes between the entire process of a single driving task and preset driving scenarios (including regular scenarios and special scenarios), and extracts three categories of indicators—safety, comfort, and efficiency—for comparison. By analyzing the differences in driving behavior between high-trust and low-trust groups under different scenarios, it can more comprehensively reflect the consistency between trust status and actual driving behavior, thereby enhancing the interpretability and empirical support of the evaluation results. Attached Figure Description

[0054] Figure 1 This is a flowchart of a method for identifying driver trust status and verifying behavior in navigation-assisted driving. Detailed Implementation

[0055] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0056] To address the shortcomings of existing methods in effectively integrating physiological signals and subjective feedback, and in verifying the quantified trust level at the behavioral level, such as... Figure 1 As shown, this invention proposes a method for identifying driver trust status and verifying behavior in assisted driving, including the following steps:

[0057] S1: Collect physiological signal data and subjective evaluation data of multiple drivers under corresponding driving tasks. The subjective evaluation data is obtained through a phased questionnaire. Calculate the driver's subjective trust score for the corresponding driving task based on the subjective evaluation data. The driving task belongs to the navigation-assisted driving process.

[0058] It should be noted that the testing equipment required for this invention mainly includes a wireless physiological signal acquisition system and a road vehicle truth acquisition system.

[0059] 1. Wireless physiological signal acquisition system

[0060] This system is used to collect physiological data such as the driver's electrocardiogram, respiration, skin conductance, and pulse. Its specific components include:

[0061] Data Acquisition Host: As the core control unit, it is responsible for receiving and processing data from various sensors.

[0062] ECG-Respiratory Module: Electrode pads are attached to specific locations on the body (such as the chest or wrist) to collect electrocardiogram signals and respiratory rate.

[0063] Skin response photoelectric pulse module: used to measure skin electrical response and pulse waveform, providing important information about the driver's emotional state.

[0064] External data synchronization module: Ensures data synchronization between various sensors and avoids data inconsistency caused by time deviation.

[0065] Analysis software: Real-time monitoring of the data collection, online filtering and calculation to obtain physiological signal data.

[0066] In practice, the first step is to correctly install electrode pads and other sensors on the driver's body. Then, the data acquisition unit and analysis software are started to synchronously collect various physiological signals from the driver. The analysis software displays the collected data in real time and performs necessary preprocessing, such as filtering and smoothing, for subsequent feature extraction and normalization.

[0067] 2. Road Vehicle True Value Acquisition System

[0068] This system is used to collect information on the vehicle's own speed and acceleration, as well as information on surrounding traffic participants. Its specific components are as follows:

[0069] GPS inertial navigation system: used to accurately record the position, speed, and acceleration / deceleration of the test vehicle.

[0070] LiDAR: Used to detect the relative position and speed of other vehicles, pedestrians, bicycles and other targets around the test vehicle with high precision.

[0071] Mobileye_EyeQ4: Specifically designed to identify environmental information such as lane lines and traffic signs, providing necessary perception support for autonomous driving functions.

[0072] Multiple cameras: distributed inside and outside the vehicle, used to record changes in the environment inside and outside the vehicle and the driver's operating behavior throughout the test.

[0073] Industrial PC: As the central processing unit, it is responsible for integrating data from all sensors and storing it on the local hard drive or uploading it to the cloud server.

[0074] Before testing, the aforementioned equipment must be installed on the actual vehicle and calibrated to ensure data accuracy. During testing, the GPS inertial navigation system continuously records the vehicle's motion, the LiDAR constantly scans the surrounding environment, the Mobileye_EyeQ4 provides lane line information in real time, and multiple cameras capture various details during the driving process. All of this data is ultimately integrated by an industrial control computer to form a complete driving task record.

[0075] In this embodiment, data collection specifically includes:

[0076] Physiological signal acquisition:

[0077] After the electrode pads and other sensors are installed on the driver, the wireless physiological signal acquisition system is activated. The system begins to synchronously collect physiological signals such as electrocardiogram, respiration, and skin conductance, and then uses analysis software to perform online filtering to obtain physiological signal data.

[0078] Vehicle operation data collection:

[0079] After installing and debugging all components of the road vehicle truth data acquisition system, start the GPS inertial navigation system, LiDAR, Mobileye_EyeQ4, and multiple cameras to begin recording vehicle speed, acceleration, information on surrounding road users, and driver behavior.

[0080] The testing process of this invention is divided into three stages: driver screening and pre-test preparation, road test data collection, and post-test subjective feedback collection. Each stage works collaboratively to simultaneously collect multimodal data and construct and verify driver trust states.

[0081] 1. First Phase: Driver Screening and Pre-Test Preparation

[0082] Before the test begins, drivers are rigorously screened to ensure their physical and mental condition meets the experimental requirements. Screening criteria include: at least 3 years of driving experience, no history of major trauma or neurological diseases, no metal implants, no physical or mental health problems that affect the acquisition of physiological signals, no use of medications that affect the autonomic nervous system, and no insomnia, excessive fatigue, or other conditions that may interfere with physiological indicators within 10 days prior to the test.

[0083] Selected drivers receive brief training covering the overall testing process, operation of the navigation-assisted driving function, how to wear the data collection device, and questionnaire completion requirements. After training, drivers complete a pre-test questionnaire, a driver behavior questionnaire, to collect basic information (such as age and gender), driving experience, daily driving style (such as conservative / aggressive), and initial attitudes and expectations regarding autonomous driving technology.

[0084] 2. Second Phase: Road Testing and Real-time Data Acquisition

[0085] Before the formal test, it was confirmed that both the wireless physiological signal acquisition system and the road vehicle truth acquisition system were in normal working order, and that all sensors were securely connected and accurately synchronized in time. The navigation system of the test vehicle was set to the predetermined test route.

[0086] After entering the test vehicle, the driver enters a 30-minute rest period. During this time, the vehicle is parked in a safe area, and the driver remains relaxed and does not perform any operations. This period is used to eliminate the driver's initial stress response upon entering the test environment, allowing physiological indicators such as heart rate and skin conductance to stabilize, ensuring the baseline reliability of subsequent data.

[0087] After the rest period, the driver drove the vehicle to the starting point of the test route. Upon arrival, the driver activated the vehicle's navigation assistance function, and the vehicle switched to autonomous driving mode, completing all test tasks along the preset route. During this process, the wireless physiological signal acquisition system and the road vehicle truth acquisition system simultaneously collected data.

[0088] The test route included multiple preset driving scenarios, such as left turns, right turns, crossing construction zones, and merging from main roads to auxiliary roads. Whenever the vehicle entered one of these scenarios, the system prompted the driver via voice to perform a phased subjective evaluation. The driver answered a set of short autonomous driving evaluation questions via voice; this voice recording was subsequently transcribed and quantified by researchers into a subjective trust score.

[0089] 3. Third stage: Collection of subjective feedback after testing

[0090] After the test was completed, all data acquisition systems were shut down and data recording was stopped. Researchers assisted the driver in removing the worn physiological signal sensor device and stored and backed up the raw data locally.

[0091] Subsequently, the driver completed a post-test questionnaire, namely the overall driving evaluation questionnaire. This questionnaire covers the overall experience of this test, including comprehensive scores for system safety, comfort, controllability, and responsiveness, as well as qualitative feedback on system performance in different scenarios, which is used to supplement the quantitative scores and support subsequent results analysis.

[0092] All data (physiological signals, vehicle operation data, phased and overall subjective scores) are archived according to driving task number and then enter the subsequent trust quantification modeling, state classification and behavior verification process.

[0093] In addition, in this embodiment, the driver's subjective trust score is calculated by integrating the phased questionnaire and the overall driving evaluation questionnaire. Specifically, each time the vehicle passes through a preset driving scenario (such as turning left, turning right, construction area, or merging from a main road and auxiliary road), the driver completes a phased subjective evaluation and obtains a trust score for that scenario. The phased scores of all scenarios are weighted according to their cognitive load or risk level in the driving task (e.g., construction area has a higher weight), and a weighted average is taken to obtain the scenario comprehensive score. This comprehensive score is further weighted and integrated with the overall driving evaluation score filled in after the test (e.g., the scenario comprehensive score accounts for 70%, and the overall score accounts for 30%) to finally obtain the driver's subjective trust score corresponding to the driving task.

[0094] S2: Feature extraction and normalization are performed on physiological signal data to obtain heart rate variability-related features and skin conductance response-related features. Grouped principal component analysis is used to reduce the dimensionality of the heart rate variability-related features and skin conductance response-related features and calculate the comprehensive heart rate variability score and comprehensive skin conductance response score corresponding to the heart rate variability-related features and skin conductance response-related features, respectively.

[0095] In this embodiment, when performing feature extraction and normalization on physiological signal data, heart rate variability-related raw features (such as average heart rate, mean RR interval, and LF / HF power ratio) and skin conductance response-related raw features (such as mean SCR waveform amplitude, mean rise time, mean recovery time, and mean SCR frequency per minute) are first extracted from the raw ECG and skin conductance signals. To eliminate the influence of differences in dimensions and orders of magnitude between different features on subsequent analysis and improve data comparability, the Min-Max normalization method is used to linearly transform the raw features. Specifically, each raw feature x is normalized according to the formula... Mapped to the interval [0,1], where x min and x max ...

[0096] The heart rate variability-related features include: average heart rate during a single driving task, mean RR interval, and the ratio of total power in the low-frequency band (LF, 0.04–0.15Hz) to total power in the high-frequency band (HF, 0.15–0.4Hz, excluding 0.15Hz) in the frequency domain analysis of heart rate variability.

[0097] The skin conductance response-related features include: the average amplitude, average rise time, average recovery time, and average frequency of SCR events per minute for all valid SCR events during a single driving task.

[0098] The effective SCR event refers to a skin conductance response event in which the change in conductance between the peak and baseline of the SCR event exceeds 0.01 μs and has a typical single-peak morphology, with a rise time between 0.5 and 5 seconds and a recovery time longer than the rise time.

[0099] The calculation methods for the heart rate variability composite score or skin conductance response composite score include:

[0100] Modeling phase: Principal component analysis is performed on the target-related feature data of multiple driving tasks to obtain multiple eigenvalues, eigenvectors corresponding to each eigenvalue, the number of retained eigenvectors (i.e., the number of principal components), and the principal component direction matrix formed by the combination of eigenvectors corresponding to the retained principal components; the principal component refers to the direction represented by the eigenvector.

[0101] The modeling phase specifically includes:

[0102] Based on the relevant feature data of multiple driving tasks, a multidimensional feature matrix is ​​formed; the covariance matrix of the multidimensional feature matrix is ​​calculated, and its eigenvalues ​​are decomposed to obtain multiple eigenvalues ​​and their corresponding eigenvectors.

[0103] For each eigenvalue, calculate the ratio of that eigenvalue to the sum of all eigenvalues ​​to obtain the variance contribution rate corresponding to that eigenvalue;

[0104] Arrange the eigenvalues ​​in descending order and accumulate the variance contribution rates corresponding to the eigenvalues ​​in turn. When the accumulated value reaches a preset threshold, the number of eigenvalues ​​or eigenvectors contained therein is the number of principal components retained.

[0105] The principal component direction matrix is ​​formed by combining the eigenvectors corresponding to the retained principal components.

[0106] Application phase: Using the principal component direction matrix obtained in the modeling phase, linear transformation is performed on the feature vectors corresponding to each principal component in the current driving task to obtain the score corresponding to each principal component; the scores corresponding to each principal component are preprocessed.

[0107] The application phase specifically includes:

[0108] The principal component direction matrix obtained in the modeling stage is used to perform a linear transformation on the feature vectors corresponding to each principal component in the current driving task to obtain the score corresponding to each principal component.

[0109] The formula for calculating the principal component score corresponding to the heart rate variability-related features is as follows:

[0110] C HRV,i =P HRV,i ·V HRV Where HRV represents heart rate variability-related characteristics;

[0111] In the formula: P HRV,i V represents the eigenvector of the i-th principal component corresponding to HRV; HRV C represents the principal component direction matrix corresponding to HRV; HRV,i This represents the score of the i-th principal component corresponding to HRV.

[0112] The formula for calculating the principal component score corresponding to the skin conductance response-related features is as follows:

[0113] C SCR,i =P SCR,i ·V SCR Among them, SCR represents skin conductance-related characteristics;

[0114] In the formula: P SCR,i V represents the feature vector corresponding to the i-th principal component of the skin conductance response-related features; SCR C represents the principal component orientation matrix corresponding to features related to skin conductance response; SCR,i This represents the score of the i-th principal component corresponding to the skin conductance response-related features.

[0115] Calculate the Pearson correlation coefficient between the scores of each principal component and the driver's subjective trust score corresponding to the same driving task; if the correlation coefficient is negative, then perform numerical inversion on the principal component score; so that the higher the adjusted scores of all principal components, the higher the corresponding driver trust level.

[0116] In the application stage of principal component analysis, this invention calculates the Pearson correlation coefficient between principal component scores and subjective trust scores, and reverses the values ​​of negatively correlated scores to ensure that the numerical direction of all composite scores is consistent with the level of trust. This approach solves the problem of inconsistent direction between physiological indicators and trust levels, making the fusion results more consistent with cognitive logic and enhancing the interpretability of the model output.

[0117] The total variance information corresponding to the target-related feature data in the current driving task is calculated based on the retained principal components, and the weight corresponding to the target-related feature data is calculated based on the total variance information.

[0118] The formula for calculating the total variance information corresponding to the heart rate variability-related features is as follows:

[0119]

[0120] The formula for calculating the total variance information corresponding to the skin electroreactivity-related features is as follows:

[0121]

[0122] In the formula: λ HRV,i λ represents the eigenvalue of the i-th retained HRV principal component;SCR,i k represents the eigenvalue of the i-th retained SCR principal component; HRV k represents the number of principal components retained corresponding to HRV. SCR Indicates the number of principal components retained corresponding to SCR; E HRV E represents the total variance information content corresponding to the dimensionality-reduced HRV; sCR This represents the total variance information content corresponding to the dimensionality-reduced SCR.

[0123] The formula for calculating the weights corresponding to the relevant feature data of the target based on the total variance information is as follows:

[0124]

[0125] In the formula, W HRV W represents the weight corresponding to HRV. SCR This indicates the weight corresponding to SCR.

[0126] The comprehensive score corresponding to the target-related feature data is calculated based on the scores of each principal component after preprocessing and the weights of the target-related feature data; the calculation formula includes:

[0127]

[0128] This invention performs feature extraction and normalization on physiological signal data, and introduces grouped principal component analysis to perform dimensionality reduction and fusion on heart rate variability-related features and skin conductance response-related features, generating comprehensive scores for heart rate variability and skin conductance response. This method avoids the interference of differences in the dimensions of different physiological indicators on the results, retains the main information features, improves the stability and interpretability of feature expression, and provides high-quality input for subsequent trust quantification.

[0129] The target-related feature data are heart rate variability-related features or skin conductance response-related features; the comprehensive score corresponding to the target-related feature data is the comprehensive score of heart rate variability and the comprehensive score of skin conductance response.

[0130] The principal components are sorted from largest to smallest according to their corresponding eigenvalues. The first principal component corresponds to the largest eigenvalue, which represents the main direction of data change. The eigenvectors corresponding to each subsequent principal component are orthogonal to the eigenvectors corresponding to all previous principal components, and their eigenvalues ​​are the largest among the remaining directions.

[0131] S3: The driver's subjective trust score, heart rate variability score, and skin conductance response score are weighted and calculated to obtain the overall trust score for the corresponding driving task; specifically:

[0132] The original scores and ratings were normalized (e.g., mapped to [0,1]) to obtain the driver's subjective trust rating, heart rate variability composite score, and skin conductance response composite score.

[0133] In this embodiment, the weight W corresponding to the driver's subjective trust score Comfort =0.399, the weight corresponding to the comprehensive score of heart rate variability is 0.318, and the weight corresponding to the skin conductance response-related features is 0.282.

[0134] The formula for calculating the overall trust score is as follows:

[0135] Comfort Index =

[0136] 100 × (0.399·Driver's Subjective Trust Score + 0.318·HRV Overall Score + 0.282·)

[0137] SCR overall score).

[0138] This invention achieves a multimodal comprehensive assessment of trust levels during assisted driving by integrating driver subjective evaluation data with physiological signal data. Compared to traditional subjective evaluation methods that rely solely on questionnaires, this invention combines the driver's genuine psychological feedback with objective physiological responses, effectively reducing the impact of individual subjective biases and improving the comprehensiveness and reliability of trust level assessment.

[0139] S4: Based on the comprehensive trust score of each driving task, a clustering algorithm is used to divide each driving task into a high trust state or a low trust state.

[0140] The comprehensive trust score based on each driving task is used to classify each driving task into a high-trust state or a low-trust state using the k-means clustering algorithm; specifically:

[0141] Randomly select k ratings from the overall trust scores of all driving tasks as the initial cluster centers;

[0142] Each data point is assigned to the nearest cluster center to form multiple clusters. The centroid of each cluster is recalculated, and the process is iterated until the centroid stabilizes or the maximum number of iterations is reached. Each data point corresponds one-to-one with the comprehensive trust score of the driving task.

[0143] The optimal number of clusters, k=2, was determined using the silhouette coefficient method and the elbow rule.

[0144] Based on the final clustering results, the cluster with the higher centroid value is defined as a high-trust state, and the cluster with the lower centroid value is defined as a low-trust state.

[0145] Each driving task is classified into a corresponding trust level state based on the cluster it belongs to.

[0146] This invention, based on the comprehensive trust score of each driving task, uses a clustering algorithm to automatically classify driving tasks into high-trust or low-trust states. The optimal number of clusters, k=2, is determined using the silhouette coefficient method and the elbow rule, achieving objective and unsupervised classification of trust states. This method eliminates the need for pre-setting thresholds, avoiding the subjectivity of manual classification and improving the scientific rigor and consistency of state recognition.

[0147] S5: Extract the vehicle operation data corresponding to the driving task, analyze the differences in driving behavior under different trust levels based on the vehicle operation data, and use statistical methods to determine whether the differences are significant in order to verify the consistency between the classified trust levels and the actual driving behavior.

[0148] The vehicle operation data includes safety, comfort, and efficiency indicators collected during the following time periods:

[0149] Vehicle operation data corresponding to the entire process of a single driving mission;

[0150] Vehicle operation data corresponding to each preset driving scenario in a single driving task. The preset driving scenarios include regular driving scenarios and special driving scenarios. Among them, regular driving scenarios include left-turn sections and right-turn sections; special driving scenarios include construction area sections and main-auxiliary road switching sections.

[0151] The security indicators include:

[0152] Minimum collision time, which represents the minimum value among the estimated collision time between the vehicle and the target ahead, is obtained by continuously monitoring the relative distance and relative speed and calculating the collision time at each moment and taking the minimum value.

[0153] Weighted arrival time (or weighted collision time) is calculated by assigning weights to the estimated collision time at each moment based on risk factors, and then calculating it through weighted averaging or weighted comprehensive calculation. It is used to reflect the overall potential collision risk level of the driving task. The risk factors include relative distance and relative speed.

[0154] In this embodiment, the system calculates the collision time estimate based on the data acquisition frequency. In this embodiment, the data acquisition frequency is 10Hz, therefore the collision time estimate is calculated every 0.1 seconds. For each time t i A risk weight is determined based on the relative distance and relative speed. (For example, the closer the distance and the greater the relative speed, the higher the weight). Weighted Time of Arrival (TTA) weighted The calculation formula is:

[0155] in:

[0156] In one implementation, risk weight This can be set using a piecewise linear function:

[0157] If the relative distance is less than 30m and the relative speed is greater than 5m / s, the weight is set to 1; otherwise, it is 0.5.

[0158] In another implementation, risk weighting It can be determined using the following formula:

[0159]

[0160] In the formula, Other variables, specifically referring to the two cars at time t. i The relative angle at any given moment;

[0161] Indicates at t i Risk weight at any given moment.

[0162] In this embodiment, the collision time estimate The calculation formula is:

[0163]

[0164] In the formula, Represents the relative distance vector, at t i The relative position between the vehicle and the target (the vehicle in front or the obstacle) at any given time; This represents the vehicle's velocity vector, i.e., the velocity vector of the vehicle at time t. i The velocity direction and magnitude at any given moment; |·|2 represents the L2 norm of the vector, i.e., the length of the vector.

[0165] Minimum relative distance, which represents the minimum relative distance between the vehicle and the target ahead;

[0166] The comfort indicators include: maximum longitudinal acceleration, minimum longitudinal acceleration, absolute mean lateral acceleration, standard deviation of heading angle, and the maximum absolute value of the rate of change of longitudinal acceleration, i.e., peak jerk.

[0167] The efficiency metrics include: average driving speed and speed standard deviation.

[0168] The S5 step is specifically as follows:

[0169] Driving tasks in high-trust and low-trust states are categorized into high-trust and low-trust groups respectively.

[0170] For both regular and special driving scenarios, the values ​​of safety, comfort, and efficiency indicators corresponding to each driving task in the two groups are extracted respectively.

[0171] For each indicator, a comparison was made between the high-confidence group and the low-confidence group: the means of the two groups were calculated, and statistical hypothesis testing methods were used to determine whether the difference was significant.

[0172] Specifically, the hypothesis testing method is used to determine whether the difference between the high-trust group and the low-trust group in a certain driving behavior indicator is statistically significant.

[0173] The independent samples t-test (suitable for data that follows a normal distribution) or the Mann-Whitney U test (suitable for non-normally distributed or small sample cases) can be used.

[0174] The significance level is usually set at α = 0.05. If the p-value obtained from the test is less than 0.05, the difference between the two groups is considered statistically significant. The p-value is a probability value calculated from the sample data using hypothesis testing methods (such as t-test, Mann-Whitney U test, etc.), representing the probability of observing the current difference or a more extreme difference given that the null hypothesis "there is no real difference between the two groups" is true.

[0175] To obtain the p-value, for example: Input: a certain indicator data (such as minimum collision time) for the high trust group and the low trust group;

[0176] Calculate: the means, standard deviation, and sample size (i.e., the number of driving tasks) for both groups;

[0177] Formula for calculating the t-statistic: In the formula, These are the index means for the high-trust group and the low-trust group, respectively. , n1 and n2 are the variances of the high-trust group and the low-trust group, respectively; n1 and n2 are the number of driving tasks in the high-trust group and the low-trust group, respectively.

[0178] Find the p-value in the t-distribution table or use software (such as Python or SPSS);

[0179] Comparing the mean differences between the two groups in "minimum collision time", if p < 0.05, it indicates that the indicator is significantly different under different trust levels, supporting the effectiveness of the trust level classification.

[0180] If, under normal driving scenarios, the mean standard deviation of longitudinal acceleration in the high-trust group is significantly smaller than that in the low-trust group, and the mean minimum collision time is significantly larger than that in the low-trust group; and under special driving scenarios, the mean minimum relative distance in the high-trust group is significantly smaller than that in the low-trust group, then the classified trust status is determined to be consistent with the actual driving behavior.

[0181] In this embodiment, under normal driving scenarios, the high-trust group exhibited more stable driving behavior: in the left-turn scenario, its minimum acceleration was significantly better than that of the low-trust group, with a difference of 0.33 m / s² in the minimum longitudinal acceleration. 2 This indicates a smoother deceleration process; in the right-turn scenario, the minimum collision time is extended by 1.00 s, and the minimum longitudinal acceleration increases by 0.26 m / s². 2 This indicates that while maintaining high traffic efficiency, a stronger safety margin is achieved. In specific scenarios, the high-trust group reduced the minimum relative distance by 1.87m and the minimum collision time by 1.87m when passing through construction areas, demonstrating more decisive following and cutting-in decisions. Although their braking action was faster and acceleration fluctuations were larger, no safety risks were posed, and subjective evaluations still showed a higher level of trust. The results show that the high-trust group prioritizes comfort and safety in conventional scenarios and demonstrates greater driving initiative and environmental adaptability in specific scenarios. The significant differences in multi-dimensional driving behavior indicators are highly consistent with the trust state grouping, verifying that the trust state classification in this invention has good external behavioral interpretability and practical distinguishing ability.

[0182] This invention further extracts vehicle operation data corresponding to driving tasks, analyzes the differences in driving behavior under different trust levels, and determines the significance of these differences through statistical hypothesis testing, thus achieving external behavioral verification of the classified trust levels. This verification mechanism establishes a correlation between psychological state and actual driving behavior, enhancing the credibility and practicality of the trust assessment results.

[0183] Furthermore, in vehicle operation data analysis, this invention distinguishes between the entire duration of a single driving task and preset driving scenarios (including regular and special scenarios), and extracts three categories of indicators—safety, comfort, and efficiency—for comparison. By analyzing the differences in driving behavior between high-trust and low-trust groups under different scenarios, it can more comprehensively reflect the consistency between trust status and actual driving behavior, enhancing the interpretability and empirical support of the evaluation results.

[0184] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0185] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0186] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0187] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A method for identifying and verifying driver trust status and behavior in assisted driving, characterized in that, Including the following steps: S1: Collect physiological signal data and subjective evaluation data of multiple drivers under corresponding driving tasks. The subjective evaluation data is obtained through a phased questionnaire. Calculate the driver's subjective trust score corresponding to the driving task based on the subjective evaluation data. The driving task described belongs to the navigation-assisted driving process; S2: Feature extraction and normalization are performed on physiological signal data to obtain heart rate variability-related features and skin conductance response-related features. Grouped principal component analysis is used to reduce the dimensionality of the heart rate variability-related features and skin conductance response-related features and calculate the comprehensive heart rate variability score and comprehensive skin conductance response score corresponding to the heart rate variability-related features and skin conductance response-related features, respectively. S3: The driver's subjective trust score, the comprehensive score of heart rate variability, and the comprehensive score of skin conductance response are weighted and calculated to obtain the comprehensive trust score for the corresponding driving task. S4: Based on the comprehensive trust score of each driving task, a clustering algorithm is used to divide each driving task into a high trust state or a low trust state. S5: Extract the vehicle operation data corresponding to the driving task, analyze the differences in driving behavior under different trust levels based on the vehicle operation data, and use statistical methods to determine whether the differences are significant in order to verify the consistency between the divided trust levels and the actual driving behavior. The vehicle operation data includes safety, comfort, and efficiency indicators collected during the following time periods: Vehicle operation data corresponding to the entire process of a single driving mission; Vehicle operation data corresponding to each preset driving scenario in a single driving task. The preset driving scenarios include regular driving scenarios and special driving scenarios. Among them, regular driving scenarios include left-turn sections and right-turn sections; special driving scenarios include construction area sections and main-auxiliary road switching sections. The security indicators include: Minimum collision time, which represents the minimum value among the estimated collision time between the vehicle and the target ahead, is obtained by continuously monitoring the relative distance and relative speed and calculating the collision time at each moment and taking the minimum value. Weighted arrival time is calculated by assigning weights to the estimated collision time at each moment based on risk factors, and then calculating it through weighted averaging or weighted comprehensive; it is used to reflect the overall potential collision risk level of the driving task; the risk factors include relative distance and relative speed. Minimum relative distance, which represents the minimum relative distance between the vehicle and the target ahead; The comfort indicators include: maximum longitudinal acceleration, minimum longitudinal acceleration, absolute mean lateral acceleration, standard deviation of heading angle, and the maximum absolute value of the rate of change of longitudinal acceleration, i.e., peak acceleration. The efficiency metrics include: average driving speed and speed standard deviation.

2. The method for identifying and verifying driver trust status and behavior in assisted driving according to claim 1, characterized in that, The heart rate variability-related features include: average heart rate during a single driving task, mean RR interval, and the ratio of total power in the low-frequency band to total power in the high-frequency band in the frequency domain analysis of heart rate variability; The skin conductance response-related features include: the average amplitude, average rise time, average recovery time, and average frequency of SCR events per minute for all valid SCR events during a single driving task. The effective SCR event is defined as a skin conductance response event in which the change in conductance between the peak and baseline of the SCR event exceeds the threshold and has a typical single-peak morphology, with a rise time between 0.5 and 5 seconds and a recovery time longer than the rise time.

3. The method for identifying and verifying driver trust status and behavior in assisted driving according to claim 1, characterized in that, The calculation methods for the heart rate variability composite score or skin conductance response composite score include: Modeling phase: Principal component analysis is performed on the target-related feature data of multiple driving tasks to obtain multiple eigenvalues, eigenvectors corresponding to each eigenvalue, the number of retained eigenvectors (i.e., the number of principal components), and the principal component direction matrix formed by the combination of eigenvectors corresponding to the retained principal components; the principal component refers to the direction represented by the eigenvector. Application phase: Using the principal component direction matrix obtained in the modeling phase, linear transformation is performed on the feature vectors corresponding to each principal component in the current driving task to obtain the score corresponding to each principal component; the scores corresponding to each principal component are preprocessed. The total variance information corresponding to the target-related feature data in the current driving task is calculated based on the retained principal components, and the weight corresponding to the target-related feature data is calculated based on the total variance information. The comprehensive score corresponding to the target-related feature data is calculated based on the scores of each principal component after preprocessing and the weights of the target-related feature data. The target-related feature data are heart rate variability-related features or skin conductance response-related features; the comprehensive score corresponding to the target-related feature data is the comprehensive score of heart rate variability and the comprehensive score of skin conductance response.

4. The method for identifying and verifying driver trust status and behavior in assisted driving according to claim 3, characterized in that, The modeling phase specifically includes: Based on the relevant feature data of multiple driving tasks, a multidimensional feature matrix is ​​formed; the covariance matrix of the multidimensional feature matrix is ​​calculated, and its eigenvalues ​​are decomposed to obtain multiple eigenvalues ​​and their corresponding eigenvectors. For each eigenvalue, calculate the ratio of that eigenvalue to the sum of all eigenvalues ​​to obtain the variance contribution rate corresponding to that eigenvalue; Arrange the eigenvalues ​​in descending order and accumulate the variance contribution rates corresponding to the eigenvalues ​​in turn. When the accumulated value reaches a preset threshold, the number of eigenvalues ​​or eigenvectors contained therein is the number of principal components retained. The principal component direction matrix is ​​formed by combining the eigenvectors corresponding to the retained principal components.

5. The method for identifying and verifying driver trust status and behavior in assisted driving according to claim 3, characterized in that, The application phase specifically includes: The principal component direction matrix obtained in the modeling stage is used to perform a linear transformation on the feature vectors corresponding to each principal component in the current driving task to obtain the score corresponding to each principal component. Calculate the Pearson correlation coefficient between the scores of each principal component and the driver's subjective trust score corresponding to the same driving task; if the correlation coefficient is negative, then perform numerical inversion on the principal component score; so that the higher the adjusted scores of all principal components, the higher the corresponding driver trust level.

6. A method for identifying and verifying driver trust status and behavior in assisted driving according to any one of claims 3 to 5, characterized in that, The principal components are sorted from largest to smallest according to their corresponding eigenvalues. The first principal component corresponds to the largest eigenvalue, which represents the main direction of data change. The eigenvectors corresponding to each subsequent principal component are orthogonal to the eigenvectors corresponding to all previous principal components, and their eigenvalues ​​are the largest among the remaining directions.

7. The method for identifying and verifying driver trust status and behavior in assisted driving according to claim 1, characterized in that, The comprehensive trust score based on each driving task uses a clustering algorithm to classify each driving task into a high-trust state or a low-trust state; specifically: Randomly select 10 scores from the overall trust scores of all driving tasks as the initial cluster centers; Each data point is assigned to the nearest cluster center to form multiple clusters. The centroid of each cluster is recalculated, and the process is iterated until the centroid stabilizes or the maximum number of iterations is reached. Each data point corresponds one-to-one with the comprehensive trust score of the driving task. The optimal number of clusters, k=2, was determined using the silhouette coefficient method and the elbow rule. Based on the final clustering results, the cluster with the higher centroid value is defined as a high-trust state, and the cluster with the lower centroid value is defined as a low-trust state. Each driving task is classified into a corresponding trust level state based on the cluster it belongs to.

8. The method for identifying and verifying driver trust status and behavior in assisted driving according to claim 7, characterized in that, The S5 step is specifically as follows: Driving tasks in high-trust and low-trust states are categorized into high-trust and low-trust groups respectively. For both regular and special driving scenarios, the values ​​of safety, comfort, and efficiency indicators corresponding to each driving task in the two groups are extracted respectively. For each indicator, a comparison was made between the high-confidence group and the low-confidence group: the means of the two groups were calculated, and statistical hypothesis testing methods were used to determine whether the difference was significant. In normal driving scenarios, the mean standard deviation of longitudinal acceleration in the high-trust group is significantly smaller than that in the low-trust group, and the mean minimum collision time is significantly larger than that in the low-trust group. In specific driving scenarios, if the mean minimum relative distance of the high trust group is significantly smaller than that of the low trust group, then the trust status is determined to be consistent with the actual driving behavior.

Images