Driver trust degree prediction method and system considering environmental parameters and perceptual risks

Abstract

The invention discloses a driver trust degree prediction method and system considering environmental parameters and perceptual risks, and the method comprises the steps: building a takeover scene feature set through collecting actual environmental data, building a plurality of automatic driving takeover scenes based on the takeover scene feature set, carrying out the automatic driving takeover simulation test, and obtaining a driver trust degree prediction result under different automatic driving takeover scenes. Taking over multi-modal data in the process; performing feature extraction on the multi-modal data, and quantifying subjective feeling features; performing feature screening by calculating correlation between subjective feeling features and objective feature sets and between subjective feeling features and takeover scene feature sets, and constructing and training an LSTM algorithm model group through a plurality of feature sets; according to the method, feature extraction is performed on environment data in an actual takeover scene through a correlation analysis algorithm and a clustering algorithm, a highly realistic simulation environment is constructed, multi-modal data of a driver is acquired, multi-dimensional data support is provided for an LSTM algorithm model group, and the reliability of a prediction model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving, and particularly to a method and system for predicting driver trust considering environmental parameters and perceived risks. Background Art

[0002] With the continuous progress of intelligent driving technology, intelligent driving systems have demonstrated increasingly powerful autonomous driving capabilities. However, in the foreseeable future, drivers still need to be ready to take over the vehicle at any time to ensure driving safety, which is both to make up for possible limitations in technology and a manifestation of drivers' active responsibility for safe driving. Therefore, drivers' vigilance and engagement still play a crucial role in the current autonomous driving experience and have a decisive impact on driving safety.

[0003] In the mode of human-machine co-driving, the driver's trust in the intelligent driving system is a key factor. The trust not only determines whether the driver is willing to rely on the intelligent driving system for driving assistance but also affects the driver's expectations and judgments of the system's response in emergency situations. The driver's trust is closely related to the perceived risk. The driver's perception of risk in different environments will affect their acceptance and trust in the autonomous driving system.

[0004] To improve the safety and comfort of autonomous driving systems, it is necessary to deeply understand the changes in driver trust and their perceived risks in different scenarios. Traditional driver trust analysis methods often only consider vehicle characteristics or driver characteristics as variables. However, in actual driving, driver trust and perceived risk are affected by multiple factors, including environmental parameters, vehicle performance, driver personal characteristics, and specific driving scenarios. Especially in different takeover scenarios, the driver's perceived risk and trust will show significant differences. For example, in a complex traffic environment, the driver may doubt the capabilities of the intelligent driving system, resulting in a decrease in trust. In a relatively simple environment, the driver may trust the system more. Therefore, simply considering vehicle or driver characteristics is not enough, and it is also necessary to comprehensively consider environmental parameters and the driver's subjective perceived risk.

[0005] In view of the fact that the human-machine co-driving mode will dominate the mainstream of transportation for a long time in the future, the present invention proposes a method and system for predicting driver trust considering environmental parameters and perceived risks to solve the above problems. Summary of the Invention

[0006] In view of the deficiencies of the above-mentioned prior art, the present invention provides a method and system for predicting driver trust considering environmental parameters and perceived risks, so as to solve the problems of insufficient prediction accuracy and inability to accurately reflect the changes in driver trust in different scenarios caused by only using vehicle characteristics or driver characteristics as variables in the prior art.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] In the first aspect, the present invention provides a method for predicting driver trust considering environmental parameters and perceived risks, including:

[0009] S1: Obtain environmental data of an autonomous vehicle in different takeover scenarios;

[0010] S2: Based on the correlation analysis algorithm and the clustering algorithm, extract features from the environmental data in different takeover scenarios, and construct a takeover scenario feature set;

[0011] S3: Based on the takeover scenario feature set, construct multiple autonomous driving takeover scenarios to conduct autonomous driving takeover simulation tests, and obtain multi-modal data during the whole process of the autonomous driving system switching working modes in different autonomous driving takeover scenarios;

[0012] The multi-modal data includes subjective feeling data, objective index data, and fixed factor data;

[0013] The subjective feeling data includes subjective trust data and subjective perceived risk data;

[0014] S4: Perform data preprocessing on the multi-modal data, and extract features from the preprocessed multi-modal data to construct a subjective feature set, an objective feature set, and a fixed feature set;

[0015] S5: Based on the subjective feature set and the objective feature set, use the random forest algorithm to screen out the key objective features strongly correlated with the subjective feeling features from the objective feature set, and construct a key objective feature set;

[0016] The key objective feature set includes a first objective feature set and a second objective feature set;

[0017] S6: Based on the subjective feature set and the key objective feature set, perform weighted fusion of the subjective and objective features through the subjective-objective combined weighting method to quantitatively calculate the perceived risk and trust of the driver;

[0018] S7: Based on the takeover scenario feature set and the quantified perceived risk and trust, use the correlation analysis algorithm to screen out the key scenario features strongly correlated with the quantified perceived risk and trust, and construct a key scenario feature set;

[0019] The key scenario feature set includes a first scenario feature set and a second scenario feature set;

[0020] S8: Construct an LSTM algorithm model group, and use the key scenario feature set, the fixed feature set, and the quantified perceived risk and trust level to train the LSTM algorithm model group;

[0021] The LSTM algorithm model group includes a perceived risk prediction model and a trust level prediction model.

[0022] Optionally, it further includes S9: Obtain the current environment data of the autonomous vehicle and input it into the trained LSTM algorithm model group to obtain a perceived risk prediction value and a trust level prediction value, and feedback the perceived risk prediction value and the trust level prediction value to the autonomous driving system.

[0023] Optionally, S2 specifically includes the following steps:

[0024] S201: Perform data preprocessing on the environment data;

[0025] The environment data includes traffic flow, road type, and weather conditions;

[0026] S202: Based on the correlation analysis algorithm, calculate the correlation coefficients between the preprocessed environment data, and screen the environment data through a preset screening threshold;

[0027] S203: Cluster the screened environment data based on the K-means algorithm to form multiple environment data clusters;

[0028] Each environment data cluster corresponds to a specific takeover scenario;

[0029] S204: Calculate the clustering center points of each environment data cluster, extract the key features of each clustering center point, and construct a takeover scenario feature set.

[0030] Optionally, S3 specifically includes the following steps:

[0031] Based on the takeover scenario feature set, construct an autonomous driving takeover simulation test scenario;

[0032] Integrate the simulation test scenario into the driving simulation platform, and construct a multimodal data acquisition device;

[0033] Run the driving simulation platform to conduct an autonomous driving takeover simulation test, and collect multimodal data during the whole process of the autonomous driving system switching working modes under different autonomous driving takeover scenarios through the multimodal data acquisition device;

[0034] The fixed factor data includes driver parameters and vehicle parameters;

[0035] The objective index data includes physiological factor data and non-driving task data.

[0036] Optionally, S5 specifically includes the following steps:

[0037] Using the objective features in the objective feature set as input parameters, and using the perceived risk feature and trust feature as output parameters respectively, construct a risk feature data set and a trust feature data set;

[0038] Divide the risk feature data set into a risk training set and a risk test set;

[0039] Divide the trust feature data set into a trust training set and a trust test set;

[0040] Construct a first random forest model, train and test the first random forest model based on the risk feature data set, to screen out the objective features strongly related to the perceived risk feature from the objective feature set, and construct a first objective feature set;

[0041] Construct a second random forest model, train and test the second random forest model based on the trust feature data set, to screen out the objective features strongly related to the trust feature from the objective feature set, and construct a second objective feature set;

[0042] The steps of constructing the first random forest model, training and testing the first random forest model based on the risk feature data set, to screen out the objective features strongly related to the perceived risk feature from the objective feature set, and constructing a first objective feature set are as follows:

[0043] Construct a first random forest model, input the risk training set into the random forest model, and perform random sampling on the risk training set to generate an initial random forest composed of multiple decision trees;

[0044] Calculate the Gini importance of each physiological feature in the random forest, sort them from high to low according to the scores, and according to the preset first score threshold, screen out the first objective features strongly related to the perceived risk feature, and construct a first objective feature set;

[0045] Based on the first objective feature set, retrain the first random forest model, and use the risk test set to verify the performance of the retrained first random forest model, to ensure that the screening process does not damage the performance of the model.

[0046] Optionally, S6 specifically includes the following steps:

[0047] Calculate the subjective weights of the first objective features in the first objective feature set based on the analytic hierarchy process;

[0048] Calculate the objective weight of the first objective feature in the first objective feature set based on the CRITIC method;

[0049] Based on the range maximization method, fuse the subjective and objective weights of each first objective feature to obtain the comprehensive weight corresponding to each first objective feature;

[0050] Perform a weighted operation on the first objective feature in the first objective feature set to obtain the objective perceived risk feature;

[0051] Fuse the objective perceived risk feature and the subjective perceived risk feature in the subjective feature set to obtain the quantified perceived risk;

[0052] Calculate the subjective weight of the second objective feature in the second objective feature set based on the analytic hierarchy process;

[0053] Calculate the objective weight of the second objective feature in the second objective feature set based on the CRITIC method;

[0054] Based on the range maximization method, fuse the subjective and objective weights of each second objective feature to obtain the comprehensive weight corresponding to each second objective feature;

[0055] Perform a weighted operation on the second objective feature in the second objective feature set to obtain the objective trustworthiness feature;

[0056] Fuse the objective trustworthiness feature and the subjective trustworthiness feature in the subjective feature set to obtain the quantified trustworthiness.

[0057] Optionally, step S7 specifically includes the following steps:

[0058] Based on the correlation analysis algorithm, calculate the correlation coefficient between the quantified perceived risk and each scenario feature in the takeover scenario feature set;

[0059] Based on a preset first correlation threshold, screen out the scenario features whose correlation coefficients exceed the first correlation threshold, and perform a significance judgment on the screened scenario features through the analysis of variance method to obtain the scenario features strongly correlated with the quantified perceived risk, and construct the first scenario feature set;

[0060] Determine the influence nature of different scenario features in the first scenario feature set on the perceived risk;

[0061] Based on the correlation analysis algorithm, calculate the correlation coefficient between the quantified trustworthiness and each scenario feature in the takeover scenario feature set;

[0062] Based on a preset second correlation threshold, screen out the scenario features whose correlation coefficients exceed the second correlation threshold, and perform significance judgment on the screened scenario features through the analysis of variance method to obtain the scenario features strongly correlated with the quantified trust degree, and construct a second scenario feature set;

[0063] Determine the influence nature of different scenario features in the second scenario feature set on the trust degree.

[0064] Optionally, the S7 specifically includes the following steps;

[0065] Construct a perceived risk prediction model based on the LSTM algorithm;

[0066] Input the first scenario feature set, the fixed feature set, and the quantified perceived risk into the perceived risk prediction model for training to obtain the trained perceived risk prediction model;

[0067] Construct a trust degree prediction model based on the LSTM algorithm;

[0068] Input the second scenario feature set, the fixed feature set, and the quantified perceived risk and trust degree into the trust degree prediction model for training to obtain the trained trust degree prediction model;

[0069] The functional expression of the LSTM algorithm is:

[0070] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i )

[0071] f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f )

[0072] c t =f t c t-1 +i t tanh(W xc x t +W hc g t-1 +b c )

[0073] o t =σ(Wxo x t +W ho h t-1 +W co c t +b o )

[0074] h t =o t tanh(c t )

[0075]

[0076] In the formula, x t is the input variable at the current time step; σ represents the sigmoid function; i t is the input gate; f t is the forget gate; c t is the cell state; o t is the output gate; h t is the hidden state; is the output layer; tanh is the hyperbolic tangent activation function; W xi ,W xf ,W xc ,W xo ,W y are the input weights of the input gate, forget gate, cell state, output gate and output layer respectively; W hi ,W hf ,W hc ,W ho are the recurrent weight matrices of the input gate, forget gate, cell state and output gate respectively; W ci ,W cf ,W co are the cell state weight matrices of the input gate, forget gate and output gate respectively; b i ,b f ,b c ,b o ,b y are the bias terms of the input gate, forget gate, cell state, output gate and output layer respectively.

[0077] In a second aspect, the present invention provides a driver trust prediction system considering environmental parameters and perceived risks, comprising:

[0078] A data acquisition module for acquiring environmental data of an autonomous vehicle in different takeover scenarios;

[0079] A feature extraction module for extracting features from environmental data in different takeover scenarios based on the correlation coefficient algorithm and the clustering algorithm, and constructing a takeover scenario feature set;

[0080] A scenario simulation module is used to construct multiple automatic driving takeover scenarios based on a takeover scenario feature set for automatic driving takeover simulation tests, and obtain multimodal data during the whole process of the automatic driving system switching working modes under different automatic driving takeover scenarios;

[0081] The multimodal data includes subjective feeling data, objective index data, and fixed factor data;

[0082] The subjective feeling data includes subjective trust degree data and subjective perceived risk data;

[0083] A data processing module is used to perform data preprocessing on the multimodal data, extract features from the preprocessed multimodal data to construct a subjective feature set, an objective feature set, and a fixed feature set;

[0084] An objective feature module is used to screen out key objective features strongly correlated with subjective feeling features from the objective feature set based on the subjective feature set and the objective feature set using the random forest algorithm, and construct a key objective feature set;

[0085] The key objective feature set includes a first objective feature set and a second objective feature set;

[0086] A quantization calculation module is used to perform weighted fusion of the subjective and objective features based on the subjective feature set and the key objective feature set through the subjective-objective combined weighting method to quantitatively calculate the perceived risk and trust degree of the driver;

[0087] A scenario feature module is used to screen out key scenario features strongly correlated with the quantitatively calculated perceived risk and trust degree based on the takeover scenario feature set and the quantitatively calculated perceived risk and trust degree using the correlation analysis algorithm, and construct a key scenario feature set;

[0088] The key scenario feature set includes a first scenario feature set and a second scenario feature set;

[0089] A model construction module is used to construct a group of LSTM algorithm models, and train the group of LSTM algorithm models using the key scenario feature set, the fixed feature set, and the quantitatively calculated perceived risk and trust degree;

[0090] The group of LSTM algorithm models includes a perceived risk prediction model and a trust degree prediction model.

[0091] Optionally, it further includes an information feedback module, which is used to obtain the current environment data of the automatic driving vehicle and input it into the trained group of LSTM algorithm models to obtain a perceived risk prediction value and a trust degree prediction value, and feedback the perceived risk prediction value and the trust degree prediction value to the automatic driving system.

[0092] The beneficial effects brought by the embodiments provided by the present invention include:

[0093] The present invention extracts features from environmental data in the actual takeover scenario through a variable correlation analysis algorithm and a clustering algorithm, constructs a highly realistic simulation environment to collect multi-modal data of the driver, provides multi-dimensional data support for the LSTM algorithm model group, and significantly improves the reliability of the prediction model.

[0094] The present invention calculates the correlation between subjective perception features and the objective feature set and the takeover scenario feature set respectively, so as to screen out the key features strongly correlated with the subjective perception features from the objective feature set and the takeover scenario feature set, improve the accuracy of predicting the driver's trust degree and perceived risk, as well as the adaptability and generalization ability of the model, and provide an accurate decision-making basis for optimizing the human-computer interaction of the autonomous driving system.

[0095] The present invention uses the LSTM algorithm model group to train based on key factors and quantified subjective perception information, mines the complex non-linear relationship between trust degree and perceived risk and multi-modal data, realizes dynamic real-time prediction, and the autonomous driving system adjusts the intervention degree according to the predicted value of the trust degree, provides auxiliary support for drivers with high trust degrees, and adopts conservative strategies for drivers with low trust degrees, thereby improving the safety performance of autonomous driving vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application, and those of ordinary skill in the art can also obtain other embodiments according to these drawings.

[0097] Figure 1 The flowchart of a method for predicting driver trust degree considering environmental parameters and perceived risk provided by an embodiment of the present specification is shown;

[0098] Figure 2 The flowchart of quantifying driver trust degree and perceived risk provided by an embodiment of the present specification is shown;

[0099] Figure 3 The flowchart of constructing an LSTM algorithm model group provided by an embodiment of the present specification is shown;

[0100] Figure 4 The structural block diagram of a system for predicting driver trust degree considering environmental parameters and perceived risk provided by an embodiment of the present specification is shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0101] The features and exemplary embodiments of various aspects of the present invention will be described in detail below. In the following detailed description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced without some of these specific details. The following description of the embodiments is only intended to provide a better understanding of the present invention by showing examples of the present invention.

[0102] In the description of the present invention, it should be noted that, unless otherwise clearly defined and limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected, or indirectly connected through an intermediate medium, and it may be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0103] Embodiment 1

[0104] This embodiment provides a method for predicting driver trust considering environmental parameters and perceived risks, including:

[0105] S1: Obtain the environmental data of the autonomous vehicle in different takeover scenarios;

[0106] Specifically, during the actual driving process of the autonomous vehicle, when a key event triggering a takeover request by the autonomous driving system occurs, collect the environmental data of the current takeover scenario; wherein, the environmental data includes but is not limited to traffic flow, road type, weather condition, conflict event, and surrounding vehicles.

[0107] S2: Based on the correlation analysis algorithm and the clustering algorithm, extract the features of the environmental data in different takeover scenarios, and construct a takeover scenario feature set;

[0108] Exemplarily, S2 specifically includes the following steps:

[0109] S201: Perform data cleaning and standardization processing on the environmental data;

[0110] S202: Based on the correlation analysis algorithm, calculate the correlation coefficients between the preprocessed environmental data, and screen the environmental data through a preset screening threshold;

[0111] In this embodiment, the correlation analysis algorithm is the Pearson correlation coefficient algorithm or the Spearman rank correlation coefficient algorithm, which can be differentially selected according to the variable type in the environmental data; the Pearson correlation coefficient is applicable to continuous normally distributed data; the Spearman correlation coefficient is applicable to rank data or non-normally distributed data.

[0112] Specifically, when the correlation analysis algorithm uses the Pearson correlation coefficient algorithm, its function expression is:

[0113]

[0114] In the formula, r is the Pearson correlation coefficient; n is the total sample size; X i is the value of the i-th X variable, and Y i is the value of the i-th Y variable. and are the means of the X variable values and the Y variable values respectively; among them, X and Y are two different types of variables among traffic flow, road type, weather condition, conflict event, and surrounding vehicles.

[0115] Specifically, when the correlation analysis algorithm uses the Spearman rank correlation coefficient algorithm, its function expression is as follows:

[0116]

[0117] d i = rank(X i ) - rank(Y i )

[0118] In the formula, ρ is the Spearman correlation coefficient; d i is the difference in ranks between the value of the i-th X variable and the value of the i-th Y variable; rank(X i ) is the rank of the value of the i-th X variable; rank(Y i ) is the rank of the value of the i-th Y variable.

[0119] Among them, according to the actual requirements, a screening threshold is set, and variable pairs with a correlation coefficient exceeding the screening threshold are selected from the environmental data. If the screening threshold is set to 0.7, when |r|≥0.7 or |ρ|≥0.7, then retain the variable pair; when |r|<0.7 or |ρ|<0.7, then eliminate the variable pair.

[0120] S203: Cluster the screened environmental data based on the K-means algorithm to form multiple environmental data clusters; among them, each environmental data cluster corresponds to a specific takeover scenario.

[0121] S204: Calculate the clustering center points of each environmental data cluster to extract the key features of each clustering center point and construct a takeover scenario feature set.

[0122] S3: Based on the takeover scenario feature set, construct multiple autonomous driving takeover scenarios to conduct autonomous driving takeover simulation tests, and obtain multimodal data during the entire process of the autonomous driving system switching working modes under different autonomous driving takeover scenarios.

[0123] Exemplarily, S3 specifically includes the following steps:

[0124] S301: Based on the takeover scenario feature set, construct an autonomous driving takeover simulation test scenario;

[0125] S302: Integrate the simulation test scenario into the driving simulation platform and construct a multi-modal data acquisition device;

[0126] Specifically, the multi-modal data acquisition device includes a physiological data acquisition unit, a subjective feeling acquisition unit, a driver information acquisition unit, a non-driving task evaluation unit, a simulated cockpit, and multiple electronic display devices;

[0127] The simulated cockpit is provided with a cockpit controller for simulating driving and is deployed with driving simulation software; the driving simulation software is used to obtain vehicle parameters; the physiological data acquisition unit is used to collect the physiological factor data of the driver; the subjective feeling acquisition unit is used to collect the subjective trust data and subjective perceived risk data of the driver under different traffic takeover scenarios; the driver information acquisition unit is used to collect driver parameters; the non-driving task acquisition unit is used to collect the non-driving task data of the driver;

[0128] In some embodiments, when the driver starts to simulate driving by starting the driving simulation software in the simulated cockpit, the fixed parameters of the vehicle are obtained through the autonomous driving system;

[0129] In some embodiments, the subjective feeling acquisition unit is a subjectively designed subjective feeling questionnaire, and the subjective evaluation questionnaire includes a perceived risk questionnaire and a trust questionnaire, which are used to obtain subjective perceived risk data and subjective trust data respectively;

[0130] The driver information acquisition unit is a subjectively designed driver information questionnaire, and the information that needs to be filled in the driver information questionnaire includes the driver's age, gender, driving style, and driving experience.

[0131] Among them, both the subjective evaluation questionnaire and the driver information questionnaire have been pre-configured into the interactive terminal in the cockpit controller. By the driver filling in the subjective evaluation questionnaire and the driver information questionnaire on the interactive terminal, the corresponding subjective feeling data and driver parameters can be obtained.

[0132] In some embodiments, the physiological data acquisition unit is a wearable physiological monitoring device; when the driver enters the simulated cockpit, the physiological data acquisition unit continuously monitors and records the physiological data of the driver in a direct contact or non-contact manner until the simulation driving ends, so as to capture the physiological state changes of the driver during the entire driving process;

[0133] Among them, the wearable physiological monitoring device includes a surface electromyograph, an electrocardiograph, a skin conductance measurement device, a respiration monitoring device, and an eye tracker.

[0134] In this embodiment, the non-driving task acquisition unit is a pre-designed non-driving task and is pre-configured in the interactive terminal of the cockpit controller. During the simulated driving process, the automatic driving system will trigger the non-driving tasks preset in the interactive terminal to require the driver to perform some tasks unrelated to driving, including but not limited to visual search tasks (such as target character recognition in a dynamic letter matrix), game tasks (such as Tetris), etc. After the non-driving task is completed, the corresponding non-driving task score will be automatically calculated.

[0135] S303: Run the driving simulation platform to conduct an automatic driving takeover simulation test, and collect multi-modal data during the whole process of the automatic driving system switching working modes in different automatic driving takeover scenarios through the multi-modal data acquisition device;

[0136] Specifically, the whole process of the vehicle automatic driving system switching to the working mode includes the process of switching from the automatic driving mode to the manual driving mode and the process of switching from the manual driving mode to the automatic driving mode;

[0137] Among them, the multi-modal data includes subjective feeling data, objective index data, and fixed factor data;

[0138] In some embodiments, the subjective feeling data includes subjective trust data and subjective perceived risk data;

[0139] In some embodiments, the objective index data includes physiological factor data and non-driving task data; the physiological factor data includes eye movement data, electrocardiogram data, skin conductance data, electromyogram data, and respiration data;

[0140] In some embodiments, the fixed factor data includes driver parameters and vehicle parameters.

[0141] S4: Perform data preprocessing on the multi-modal data, and extract features from the preprocessed multi-modal data to construct a subjective feature set, an objective feature set, and a fixed feature set;

[0142] Exemplarily, S4 specifically includes the following steps:

[0143] S401: Perform data preprocessing on the multi-modal data, and align and associate the preprocessed multi-modal data according to the time stamp;

[0144] Specifically, the data preprocessing includes denoising processing, outlier removal processing, filling in missing values, and standardization processing;

[0145] In this embodiment, noise signals in multi-modal data are removed through a digital filtering algorithm, abnormal data points in obvious multi-modal data are identified and eliminated through logical judgment rules, and possible data missing situations in multi-modal data are complemented using a data interpolation algorithm to ensure data continuity and integrity; finally, the multi-modal data in different dimensions are standardized so that the multi-dimensional data are uniformly converted to the standard numerical range for subsequent comprehensive analysis;

[0146] S402: Extract relevant features of subjective perception data, physiological factor data, and fixed factor data to construct a subjective feature set, an objective feature set, and a fixed feature set;

[0147] In some embodiments, the relevant features of vehicle parameters include but are not limited to the autonomous driving style, the torque when the vehicle accelerates, the braking force when the vehicle decelerates, and the visualized vehicle interior information.

[0148] In some embodiments, the relevant features of driver parameters include but are not limited to driver age, driver gender, driving style, and driving experience;

[0149] In some embodiments, the relevant features of physiological factor data include but are not limited to the pupil diameter of the driver, the fixation duration of the region of interest, the fixation frequency of the region of interest, the heart rate variability of the driver, the heartbeat interval of the driver, the skin conductance response related to takeover events, the peak value of the skin conductance response, the root mean square value of the electromyogram signal, the activation threshold of the electromyogram signal, the respiratory frequency, and the maximum peak value of respiration;

[0150] In some embodiments, the relevant feature of non-driving task data is the non-driving task score;

[0151] In some embodiments, the relevant features of subjective evaluation data include the subjective trust score and the subjective perceived risk score;

[0152] S5: Based on the subjective feature set and the objective feature set, using the random forest algorithm, screen out the key objective features strongly related to the subjective perception features from the objective feature set and construct a key objective feature set;

[0153] Among them, the key objective feature set includes a first objective feature set and a second objective feature set;

[0154] Exemplarily, S5 specifically includes the following steps:

[0155] S501: Use the objective features in the objective feature set as input parameters, and use the perceived risk feature and the trust feature as output parameters respectively to construct a risk feature data set and a trust feature data set;

[0156] S502: Divide the risk feature dataset into a risk training set and a risk test set according to a preset ratio; divide the trust feature dataset into a trust training set and a trust test set according to a preset ratio;

[0157] S503: Construct a first random forest model, train and test the first random forest model based on the risk feature dataset, so as to screen out the objective features strongly related to the perceived risk features from the objective feature set, and construct a first objective feature set;

[0158] S504: Construct a second random forest model, train and test the second random forest model based on the trust feature dataset, so as to screen out the objective features strongly related to the trust degree features from the objective feature set, and construct a second objective feature set.

[0159] In some embodiments, S503 specifically includes the following steps:

[0160] Construct a first random forest model, input the risk training set into the random forest model, and randomly sample the risk training set to generate an initial random forest composed of multiple decision trees;

[0161] Calculate the Gini importance of each objective feature in the random forest, sort them from high to low according to the score, and according to a preset first score threshold, screen out the first objective features strongly related to the perceived risk features, and construct a first objective feature set;

[0162] Based on the first objective feature set, retrain the first random forest model, and use the risk test set to verify the performance of the retrained first random forest model, so as to ensure that the screening process does not damage the performance of the model.

[0163] In some embodiments, S504 specifically includes the following steps:

[0164] Construct a second random forest model, input the risk training set into the second random forest model, and randomly sample the trust training set to generate an initial random forest composed of multiple decision trees;

[0165] Calculate the Gini importance of each objective feature in the random forest, sort them from high to low according to the score, and according to a preset second score threshold, screen out the second objective features strongly related to the trust degree features, and construct a second objective feature set;

[0166] Based on the first objective feature set, retrain the first random forest model, and use the risk test set to verify the performance of the retrained first random forest model, so as to ensure that the screening process does not damage the performance of the model.

[0167] S6: Based on the subjective feature set and the key objective feature set, the subjective and objective features are weighted and fused through the subjective and objective combined weighting method to quantitatively calculate the perceived risk and trust of the driver;

[0168] Exemplarily, S6 specifically includes the following steps:

[0169] S601: Calculate the subjective weight of the first objective feature in the first objective feature set based on the analytic hierarchy process;

[0170] S602: Calculate the objective weight of the first objective feature in the first objective feature set based on the CRITIC method;

[0171] S603: Based on the range maximization method, weight-fuse the subjective and objective weights of each first objective feature to obtain the comprehensive weight corresponding to each first objective feature;

[0172] S604: Perform a weighted operation on the first objective feature in the first objective feature set to obtain the objective perceived risk score;

[0173] S605: Fuse the objective perceived risk score and the subjective perceived risk score to obtain the quantified perceived risk;

[0174] S606: Calculate the subjective weight of the second objective feature in the second objective feature set based on the analytic hierarchy process;

[0175] S607: Calculate the objective weight of the second objective feature in the second objective feature set based on the CRITIC method;

[0176] S608: Based on the range maximization method, weight-fuse the subjective and objective weights of each second objective feature to obtain the comprehensive weight corresponding to each second objective feature;

[0177] S609: Perform a weighted operation on the second objective feature in the second objective feature set to obtain the objective trust score;

[0178] S610: Fuse the objective trust score and the subjective trust score to obtain the quantified trust.

[0179] S7: Based on the takeover scenario feature set and the quantified perceived risk and trust, use the correlation analysis algorithm to screen out the key scenario features that are strongly correlated with the quantified perceived risk and trust, and construct a key scenario feature set;

[0180] Among them, the key scenario feature set includes a first scenario feature set and a second scenario feature set;

[0181] Exemplarily, S7 specifically includes the following steps:

[0182] S701: Calculate the correlation coefficient between the quantified perceived risk and each scenario feature in the takeover scenario feature set based on the correlation analysis algorithm.

[0183] In this embodiment, the correlation analysis algorithm is the Pearson correlation coefficient algorithm or the Spearman rank correlation coefficient algorithm, which can be differentially selected according to the variable type of each scenario feature in the takeover scenario feature set.

[0184] S702: Based on a preset first correlation threshold, screen out the scenario features whose correlation coefficients exceed the first correlation threshold, and perform a significance judgment on the screened scenario features through the analysis of variance method to obtain the scenario features strongly correlated with the quantified perceived risk, and construct the first scenario feature set.

[0185] Specifically, performing a significance judgment on the screened scenario features through the analysis of variance method specifically includes the following steps:

[0186] According to the corresponding takeover scenarios to which each scenario feature belongs, divide all the screened scenario features into different takeover scenario groups.

[0187] Calculate the total sample mean of all the screened scenario features and the group mean of each takeover scenario group.

[0188] Based on the overall mean of all the screened scenario features and the group mean of each takeover scenario group, calculate the sum of squares between groups; based on the group mean of each takeover scenario group, calculate the sum of squares within groups.

[0189] Based on the sum of squares between groups and the degrees of freedom between groups, calculate the mean square between groups to reflect the degree of difference between different takeover scenarios.

[0190] Based on the sum of squares within groups and the degrees of freedom within groups, calculate the mean square within groups to reflect the random error (noise) of the scenario features within the same takeover scenario.

[0191] Calculate the F value through the ratio of the mean square between groups and the mean square within groups, and calculate the P value based on the degrees of freedom between groups, the degrees of freedom within groups, and the F value.

[0192] Among them, the larger the F value, the more significant the difference between groups; the P value is used to judge whether the observed effect (such as the difference between groups, correlation, etc.) may be caused by random error, and when p < 0.05 is the significance criterion; in this embodiment, by quantifying the ratio of the difference between groups (the impact of different takeover scenarios on the perceived risk) and the difference within groups (the random error between features within the same takeover scenario), judge whether there is a significant difference in the impact of different takeover scenario features on the perceived risk. When the F value is much greater than 1 and the P value is less than 0.05, it indicates that the difference between groups is much greater than the random error, and different takeover scenario features have a significant impact on the quantified perceived risk.

[0193] In this embodiment, when the absolute value of the correlation coefficient of each scenario feature in the takeover scenario feature set exceeds the first correlation threshold, the scenario feature is retained; otherwise, the scenario feature is eliminated. At the same time, a significance judgment is made on the retained scenario features. When the significance p of the retained scenario feature is ≥ 0.05, the scenario feature is eliminated; when the significance p of the retained scenario feature is < 0.05, the retained scenario feature is a key scenario feature strongly correlated with the quantified perceived risk; and a first scenario feature set is constructed based on the key scenario features.

[0194] S703: Determine the nature of the influence of different scenario features in the first scenario feature set on the perceived risk;

[0195] Specifically, by comparing the mean values of the perceived risks of different scenario features in the first scenario feature set, it is judged whether the influence of different takeover scenario features on the perceived risk features is positive feedback or negative feedback;

[0196] If the scenario feature is positively correlated with the perceived risk, the original value or the standardized continuous value can be directly input into the subsequent model, because its numerical size directly corresponds to the degree of risk.

[0197] If the scenario feature is negatively correlated with the risk, the feature needs to be processed in the reverse direction to avoid confusing the learning logic of the subsequent model due to the opposite direction of the numerical value and the risk direction.

[0198] S704: Calculate the correlation coefficients between the quantified trust degree and each scenario feature in the takeover scenario feature set based on the correlation analysis algorithm;

[0199] S705: Screen out the scenario features whose correlation coefficients exceed the second correlation threshold based on the preset second correlation threshold, and perform a significance judgment on the screened scenario features through the analysis of variance method to obtain the scenario features strongly correlated with the quantified trust degree, and construct a second scenario feature set;

[0200] S706: Determine the nature of the influence of different scenario features in the second scenario feature set on the trust degree.

[0201] In this embodiment, the process of screening out the key scenario features strongly correlated with the quantified trust degree and constructing the second scenario feature set is basically the same as steps S701 - S703, and the difference is that the screened feature sets and screening directions of the two are different.

[0202] S8: Construct an LSTM algorithm model group, and train the LSTM algorithm model group by using the key scenario feature set, the fixed feature set, and the quantified perceived risk and trust degree;

[0203] Among them, the LSTM algorithm model group includes a perceived risk prediction model and a trust degree prediction model;

[0204] In this embodiment, the LSTM algorithm (Long Short-Term Memory) is a special type of Recurrent Neural Network (RNN) that can effectively capture long-term dependencies in sequential data and is widely used in fields such as natural language processing, time series prediction, and speech recognition.)

[0205] Exemplarily, S8 specifically includes the following steps;

[0206] S801: Construct a perceived risk prediction model based on the LSTM algorithm, and input the first scenario feature set, fixed feature set, and quantified perceived risk into the perceived risk prediction model for training to obtain the trained perceived risk prediction model;

[0207] Specifically, S801 specifically includes the following content:

[0208] Based on the LSTM algorithm, construct a perceived risk prediction model;

[0209] Define the LSTM network structure to include an LSTM layer and an output layer, and initialize the parameters of the perceived risk prediction model including weights and biases;

[0210] Define a loss function, including defining the mean squared error, which is used to measure the difference between the model prediction result and the actual data; among them, the loss function in this embodiment is the mean squared error function (MSE) or the cross-entropy loss function;

[0211] Update the model parameters including stochastic gradient descent through the backpropagation algorithm and optimization method, reduce the loss function, and improve the prediction accuracy of the model;

[0212] Input the first scenario feature set, fixed feature set, and quantified perceived risk into the perceived risk prediction model for model training, calculate the error value through the loss function and backpropagate to update the parameters of the perceived risk prediction model, and continuously iterate the training until reaching the predetermined number of training times or meeting the convergence condition and then stop training to obtain the trained perceived risk prediction model;

[0213] The functional expression of the LSTM algorithm is:

[0214] i t = σ(W xi x t + W hi h t-1 + W ci c t-1 + b i )

[0215] f t = σ(W xfx t +W hf h t-1 +W cf c t-1 +b f )

[0216] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c )

[0217] o t =σ(W xo x t +W ho h t-1 +W co c t +b o )

[0218] h t =o t tanh(c t )

[0219]

[0220] In the formula, x t is the input variable at the current time step; σ represents the sigmoid function; i t is the input gate, with a value range of [0, 1], which is used to control the inflow of new information and determine the weight of environmental parameters at this moment. The more important the parameter, the greater the weight. All environmental parameters and fixed parameters share the same input gate; f t is the forget gate, with a value range of [0, 1], which controls the retention degree of old information and judges whether to forget the parameters at the previous moment; c t is the cell state, which updates the cell state by first retaining the previous data, then adding the input of the current state and the bias term; o t is the output gate, with a value range of [0, 1], which controls the proportion of the current memory output to the outside world and determines how much information of the current cell state needs to be output as the hidden state; h t is the hidden state, the hidden state at the current moment, which is passed to the next moment to update the cell parameters and is also input to the output layer to obtain the prediction value at the current moment; is the output layer, the predicted output at the current moment; tanh is the hyperbolic tangent activation function; W xi ,W xf ,W xc ,Wxo and W y are respectively the input weights of the input gate, forget gate, cell state, output gate, and output layer; W hi and W hf and W hc and W ho are respectively the recurrent weight matrices of the input gate, forget gate, cell state, and output gate; W ci and W cf and W co are respectively the cell state weight matrices of the input gate, forget gate, and output gate; b i and b f and b c and b o and b y are respectively the bias terms of the input gate, forget gate, cell state, output gate, and output layer.

[0221] S802: Construct a trust prediction model based on the LSTM algorithm, and input the second scenario feature set, fixed feature set, and the quantified perceived risk and trust into the trust prediction model for training to obtain the trained trust prediction model;

[0222] In this embodiment, the training process of the trust prediction model is basically the same as that of the above-mentioned perceived risk prediction model, except that the input parameters of the two training models are different.

[0223] In some embodiments, it further includes S9: Obtain the current environmental data of the autonomous vehicle and input it into the trained LSTM algorithm model group to obtain the perceived risk prediction value and trust prediction value, and feedback the perceived risk prediction value and trust prediction value to the autonomous driving system.

[0224] Specifically, S9 specifically includes the following steps:

[0225] Obtain the current environmental data of the autonomous vehicle;

[0226] Input the current environmental data into the trained perceived risk prediction model to obtain the perceived risk prediction value;

[0227] Input the perceived risk prediction value and the previous environmental data into the trained trust prediction model to obtain the trust prediction value;

[0228] Feedback the perceived risk prediction value and trust prediction value to the autonomous driving system.

[0229] It should be clearly stated that before driving an autonomous vehicle, the driver will conduct a questionnaire survey to collect the driver's fixed parameters. At the same time, the fixed parameters of the vehicle are fixed when the vehicle leaves the factory and can be directly obtained through the autonomous driving system. Therefore, in the following, only environmental data needs to be obtained, and the perceived risk prediction value and the trust prediction value can be calculated through the environmental data.

[0230] In this embodiment, by effectively combining the perceived risk prediction value, the trust prediction value, and the real-time environmental data, the dynamic change of the driver's trust in the upcoming driving scenario can be accurately predicted and grasped. This not only helps the autonomous driving system better understand the driver's mental state in different situations but also provides an important reference basis for the real-time adjustment and optimization of the autonomous driving system. By using the trust prediction value obtained through comprehensive analysis, the autonomous driving system can provide different assisted driving strategies according to the driver's different levels of trust. For example, when the driver does not trust, the autonomous driving system selects a slower and smoother driving mode to improve the driver's subjective trust.

[0231] Embodiment 2

[0232] As Figure 4 shown, this embodiment provides a driver trust prediction system considering environmental parameters and perceived risks, including:

[0233] A data acquisition module for acquiring environmental data of the autonomous vehicle in different takeover scenarios;

[0234] A feature extraction module for extracting features from the environmental data in different takeover scenarios based on the correlation coefficient algorithm and the clustering algorithm, and constructing a takeover scenario feature set;

[0235] A scenario simulation module for constructing multiple autonomous driving takeover scenarios based on the takeover scenario feature set to conduct autonomous driving takeover simulation tests, and obtaining multi-modal data during the whole process of the autonomous driving system switching working modes in different autonomous driving takeover scenarios;

[0236] In some embodiments, the multi-modal data includes subjective feeling data, objective index data, and fixed factor data;

[0237] In some embodiments, the subjective feeling data includes subjective trust data and subjective perceived risk data;

[0238] A data processing module for preprocessing the multi-modal data, and extracting features from the preprocessed multi-modal data to construct a subjective feature set, an objective feature set, and a fixed feature set;

[0239] An objective feature module, which is used to screen out key objective features strongly related to subjective feeling features from the objective feature set based on the subjective feature set and the objective feature set by using the random forest algorithm, and construct a key objective feature set;

[0240] In some embodiments, the key objective feature set includes a first objective feature set and a second objective feature set;

[0241] A quantization calculation module, which is used to perform weighted fusion of subjective and objective features based on the subjective feature set and the key objective feature set by using the subjective-objective combined weighting method, so as to realize the quantization calculation of the driver's perceived risk and trust;

[0242] A scenario feature module, which is used to screen out key scenario features strongly related to the quantified perceived risk and trust based on the takeover scenario feature set and the quantified perceived risk and trust, and construct a key scenario feature set by using the correlation analysis algorithm;

[0243] In some embodiments, the key scenario feature set includes a first scenario feature set and a second scenario feature set;

[0244] A model construction module, which is used to construct a group of LSTM algorithm models, and train the group of LSTM algorithm models by using the key scenario feature set, the fixed feature set, and the quantified perceived risk and trust;

[0245] The group of LSTM algorithm models includes a perceived risk prediction model and a trust prediction model.

[0246] In some embodiments, it further includes an information feedback module, which is used to obtain the current environment data of the autonomous vehicle and input it into the trained group of LSTM algorithm models to obtain a perceived risk prediction value and a trust prediction value, and feedback the perceived risk prediction value and the trust prediction value to the autonomous driving system.

[0247] In some embodiments, the data acquisition module specifically includes, but is not limited to, any one or a combination of GPS, wireless communication devices, on-vehicle cameras, and lidar.

[0248] As described above, only the preferred embodiments of the present invention are provided. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to include all changes falling within the meaning and scope of the equivalent elements of the claims in the present invention.

Claims

1. A method for predicting driver trust considering environmental parameters and perceived risks, characterized in that, Including: S1: Obtain the environmental data of the autonomous vehicle in different takeover scenarios; S2: Based on the correlation analysis algorithm and the clustering algorithm, extract the features of the environmental data in different takeover scenarios, and construct a takeover scenario feature set; S3: Based on the takeover scenario feature set, construct multiple autonomous driving takeover scenarios for autonomous driving takeover simulation tests, and obtain multi-modal data during the whole process of the autonomous driving system switching working modes in different autonomous driving takeover scenarios; The multi-modal data includes subjective feeling data, objective index data and fixed factor data; The subjective feeling data includes subjective trust degree data and subjective perceived risk data; S4: Perform data preprocessing on the multi-modal data, and extract features from the preprocessed multi-modal data to construct a subjective feature set, an objective feature set and a fixed feature set; S5: Based on the subjective feature set and the objective feature set, use the random forest algorithm to screen out the key objective features strongly correlated with the subjective feeling features from the objective feature set, and construct a key objective feature set; The key objective feature set includes a first objective feature set and a second objective feature set; S6: Based on the subjective feature set and the key objective feature set, perform weighted fusion of the subjective and objective features through the subjective-objective combined weighting method to realize the quantitative calculation of the driver's perceived risk and trust degree; S7: Based on the takeover scenario feature set and the quantified perceived risk and trust degree, use the correlation analysis algorithm to screen out the key scenario features strongly correlated with the quantified perceived risk and trust degree, and construct a key scenario feature set; The key scenario feature set includes a first scenario feature set and a second scenario feature set; S8: Construct an LSTM algorithm model group, and use the key scenario feature set, the fixed feature set, and the quantified perceived risk and trust degree to train the LSTM algorithm model group; The LSTM algorithm model group includes a perceived risk prediction model and a trust degree prediction model.

2. The method according to claim 1, wherein It also includes S9: Obtain the current environmental data of the autonomous vehicle, and input it into the trained LSTM algorithm model group to obtain a perceived risk prediction value and a trust degree prediction value, and feedback the perceived risk prediction value and the trust degree prediction value to the autonomous driving system.

3. The method according to claim 1, wherein The specific steps of S2 are as follows: S201: Perform data preprocessing on the environmental data; The environmental data includes traffic flow, road type and weather condition; S202: Based on the correlation analysis algorithm, calculate the correlation coefficient between the preprocessed environmental data, and screen the environmental data through a preset screening threshold; S203: Cluster the screened environmental data based on the K-means algorithm to form multiple environmental data clusters; Each environmental data cluster corresponds to a specific takeover scenario; S204: Calculate the clustering center points of each environmental data cluster to extract the key features of each clustering center point, and construct a takeover scenario feature set.

4. The method according to claim 1, wherein The specific steps of S3 are as follows: Based on the takeover scenario feature set, construct an autonomous driving takeover simulation test scenario; Integrate the simulation test scenario into the driving simulation platform, and construct a multi-modal data acquisition device; Run the driving simulation platform to conduct autonomous driving takeover simulation tests, and collect multimodal data during the whole process of the autonomous driving system switching working modes in different autonomous driving takeover scenarios through a multimodal data acquisition device; The fixed factor data includes driver parameters and vehicle parameters; The objective index data includes physiological factor data and non-driving task data.

5. The method according to claim 1, wherein The specific steps of S5 are as follows: Use the objective features in the objective feature set as input parameters, and use the perceived risk feature and trust feature as output parameters respectively to construct a risk feature data set and a trust feature data set; Divide the risk feature data set into a risk training set and a risk test set; Divide the trust feature data set into a trust training set and a trust test set; Construct a first random forest model, train and test the first random forest model based on the risk feature data set to screen out the objective features strongly related to the perceived risk feature from the objective feature set, and construct a first objective feature set; Construct a second random forest model, train and test the second random forest model based on the trust feature data set to screen out the objective features strongly related to the trust feature from the objective feature set, and construct a second objective feature set; The specific steps of constructing the first random forest model, training and testing the first random forest model based on the risk feature data set to screen out the objective features strongly related to the perceived risk feature from the objective feature set, and constructing the first objective feature set are as follows: Construct a first random forest model, input the risk training set into the random forest model, and randomly sample the risk training set to generate an initial random forest composed of multiple decision trees; Calculate the Gini importance of each physiological feature in the random forest, sort them from high to low according to the scores, and screen out the first objective features strongly related to the perceived risk feature according to the preset first score threshold, and construct a first objective feature set; Based on the first objective feature set, retrain the first random forest model, and use the risk test set to verify the performance of the retrained first random forest model to ensure that the screening process does not damage the performance of the model.

6. The method according to claim 1, characterized in that The specific steps of S6 are as follows: Calculate the subjective weights of the first objective features in the first objective feature set based on the analytic hierarchy process; Calculate the objective weights of the first objective features in the first objective feature set based on the CRITIC method; Based on the range maximization method, fuse the subjective and objective weights of each first objective feature to obtain the comprehensive weight corresponding to each first objective feature; Perform a weighted operation on the first objective features in the first objective feature set to obtain the objective perceived risk feature; Fuse the objective perceived risk feature and the subjective perceived risk feature in the subjective feature set to obtain the quantified perceived risk; Calculate the subjective weights of the second objective features in the second objective feature set based on the analytic hierarchy process; Calculate the objective weights of the second objective features in the second objective feature set based on the CRITIC method; Based on the range maximization method, fuse the subjective and objective weights of each second objective feature to obtain the comprehensive weight corresponding to each second objective feature; Perform a weighted operation on the second objective features in the second objective feature set to obtain objective trust degree features; Fuse the objective trust degree features and the subjective trust degree features in the subjective feature set to obtain the quantified trust degree.

7. The method according to claim 1, characterized in that, The specific steps of S7 are as follows: Based on the correlation analysis algorithm, calculate the correlation coefficients between the quantified perceived risk and each scenario feature in the takeover scenario feature set; Based on a preset first correlation threshold, filter out the scenario features whose correlation coefficients exceed the first correlation threshold, and perform a significance judgment on the filtered scenario features through the analysis of variance method to obtain the scenario features strongly correlated with the quantified perceived risk, and construct the first scenario feature set; Determine the influence nature of different scenario features in the first scenario feature set on the perceived risk; Based on the correlation analysis algorithm, calculate the correlation coefficients between the quantified trust degree and each scenario feature in the takeover scenario feature set; Based on a preset second correlation threshold, filter out the scenario features whose correlation coefficients exceed the second correlation threshold, and perform a significance judgment on the filtered scenario features through the analysis of variance method to obtain the scenario features strongly correlated with the quantified trust degree, and construct the second scenario feature set; Determine the influence nature of different scenario features in the second scenario feature set on the trust degree.

8. The method according to claim 1, characterized in that, The specific steps of S8 are as follows; Construct a perceived risk prediction model based on the LSTM algorithm; Input the first scenario feature set, the fixed feature set, and the quantified perceived risk into the perceived risk prediction model for training to obtain the trained perceived risk prediction model; Construct a trust degree prediction model based on the LSTM algorithm; Input the second scenario feature set, the fixed feature set, and the quantified perceived risk and trust degree into the trust degree prediction model for training to obtain the trained trust degree prediction model; The function expression of the LSTM algorithm is: i t = σ(W xi x t + W hi h t-1 + W ci c t-1 + b i ) f t = σ(W xf x t + W hf h t-1 + W cf c t-1 + b f ) c t = f t c t-1 + i t tanh(W xc x t + W hc h t-1 + b c ) o t = σ(W xo x t + W ho h t-1 + W co c t + b o ) h t = o t tanh(c t ) where x t is the input variable at the current time step; σ represents the sigmoid function; i t is the input gate; f t is the forget gate; c t is the cell state; o t is the output gate; h t is the hidden state; is the output layer; tanh is the hyperbolic tangent activation function; W xi , W xf , W xc , W xo , W y are the input weights of the input gate, forget gate, cell state, output gate, and output layer, respectively; W hi , W hf , W hc , W ho are the recurrence weight matrices of the input gate, forget gate, cell state, and output gate, respectively; W ci , W cf , W co are the cell state weight matrices of the input gate, forget gate, and output gate, respectively; b i , b f , b c , b o , b y are the bias terms of the input gate, forget gate, cell state, output gate, and output layer, respectively.

9. A driver trust prediction system that takes into account environmental parameters and perceived risks, characterized in that, Including: A data acquisition module for acquiring the environmental data of the autonomous driving vehicle in different takeover scenarios; A feature extraction module for extracting features from the environmental data in different takeover scenarios based on the correlation coefficient algorithm and the clustering algorithm, and constructing a takeover scenario feature set; A scenario simulation module for constructing multiple autonomous driving takeover scenarios based on the takeover scenario feature set to perform autonomous driving takeover simulation tests, and obtaining multi-modal data in the whole process of the autonomous driving system switching working modes in different autonomous driving takeover scenarios; The multi-modal data includes subjective feeling data, objective index data, and fixed factor data; The subjective feeling data includes subjective trust degree data and subjective perceived risk data; A data processing module for preprocessing the multi-modal data, and extracting features from the preprocessed multi-modal data to construct a subjective feature set, an objective feature set, and a fixed feature set; An objective feature module for screening out the key objective features strongly correlated with the subjective feeling features from the objective feature set based on the subjective feature set and the objective feature set by using the random forest algorithm, and constructing a key objective feature set; The key objective feature set includes a first objective feature set and a second objective feature set; A quantization calculation module, which is used to perform weighted fusion of subjective and objective features through a subjective and objective combined weighting method based on a subjective feature set and a key objective feature set, so as to realize the quantization calculation of the perceived risk and trust degree of the driver; A scenario feature module, which is used to screen out key scenario features strongly correlated with the quantized perceived risk and trust degree by using a correlation analysis algorithm based on a takeover scenario feature set and the quantized perceived risk and trust degree, and construct a key scenario feature set; The key scenario feature set includes a first scenario feature set and a second scenario feature set; A model construction module, which is used to construct a group of LSTM algorithm models, and train the group of LSTM algorithm models by using the key scenario feature set, a fixed feature set, and the quantized perceived risk and trust degree; The group of LSTM algorithm models includes a perceived risk prediction model and a trust degree prediction model.

10. The system according to claim 9, wherein It further includes an information feedback module, which is used to obtain the current environmental data of the autonomous vehicle and input it into the trained group of LSTM algorithm models to obtain a perceived risk prediction value and a trust degree prediction value, and feedback the perceived risk prediction value and the trust degree prediction value to the autonomous driving system.

Images