Autonomous driving vehicle control method based on driving style and dynamic trust evaluation

Abstract

The present invention belongs to the field of road vehicle control systems and relates to a method for controlling an autonomous driving vehicle based on driving style and dynamic trust evaluation. The method identifies the driving style based on the driver's real-time operating behavior; introduces the FiLM-Gate style modulation mechanism to optimize the LSTM neural network, and predicts the speed and acceleration at the next moment based on the optimized LSTM neural network; generates personalized control parameters in real time based on the driving style and prediction information; at the same time, collects the driver's response characteristics to the autonomous driving behavior, and dynamically estimates the driver's trust state in the autonomous driving system based on the POMDP. The trust estimation result is used as a supplementary adjustment factor to make the optimal action decision, and the style-trust dual-channel fusion control vehicle is controlled, ultimately realizing the joint closed-loop control of "personalization + trust perception + safety optimization", comprehensively improving the reliability and user acceptance of autonomous driving vehicles, and improving the safety level of human-machine co-driving.

Description

Technical Field

[0001] The present invention belongs to the field of road vehicle control systems and relates to an autonomous driving vehicle control method based on driving style and dynamic trust evaluation. Background Art

[0002] Currently, most self-driving cars available to the public are Level 2 and Level 3 automated, allowing drivers to shift their attention from the primary task of driving, but they must still be prepared to take control of the vehicle when necessary. Research has shown that automated vehicle driving can reduce human error, reduce traffic injuries and fatalities on the road, and significantly improve road traffic safety. However, there are still many concerns about social acceptance of self-driving cars, and trust in handing over full vehicle control to the self-driving system remains low. Drivers' trust in self-driving cars affects the level of human-machine co-driving. In this co-driving, the self-driving system participates in the overall system control as a partner of the driver. The existence of trust helps establish a good collaborative relationship, improving system efficiency and driving safety.

[0003] Currently, autonomous driving technology still faces many challenges. For example, autonomous vehicles with an average driving style will not satisfy all drivers. Different drivers have different perceptions of comfort; while some drivers may prefer an aggressive driving style, others may prefer a conservative one. Therefore, it is crucial to customize the vehicle's autonomous driving functions according to the driver's actual driving style. In addition, people's acceptance of autonomous vehicles does not seem to keep pace with the growing market penetration, partly due to trust issues in the technology. Autonomous driving decision-making systems should be able to react to and change the human trust dynamic during interactions, rather than blindly following procedures. These challenges need to be effectively addressed and resolved in the process of promoting the development of automotive automation technology.

[0004] Chinese patent CN 116975671 A proposes a method for assessing driver trust in Level 3 autonomous vehicles, combining a subjective trust scale and driving information from key takeover frames to assess driver trust. Chinese patent CN119550992 A measures driver trust from both subjective and objective perspectives, quantifying subjective trust based on an autonomous driving contextual trust scale and objective trust by measuring driver physiological indicators. While current research can assess driver trust to a certain extent, it still has several shortcomings and deficiencies. First, regarding applicable scenarios, current research is mostly limited to assessing trust when the driver intends to take over the autonomous driving system, failing to consider the overall driver trust level and adjust the autonomous driving system accordingly. Second, driver trust assessments often suffer from issues such as a single metric and heavy reliance on physiological data, which interferes with driver data and subjective trust scales during collection. Furthermore, these studies often overlook the key subjective factor of driving style. Summary of the Invention

[0005] In view of the defects and shortcomings of the existing technology, the purpose of the present invention is to provide an autonomous vehicle control method based on driving style and dynamic trust assessment. During vehicle operation, this method identifies the driver's driving style based on the driver's real-time operating behavior. Through a long short-term memory network (LSTM) structure, the driving style identification results are integrated to generate or update personalized control parameters that better meet the driver's expectations in real time. At the same time, the system continuously collects the driver's response characteristics to the autonomous driving behavior and dynamically estimates the driver's trust state in the autonomous driving system based on a partially observable Markov decision process (POMDP) ​​modeling framework. The above driving style identification results are used as the basis for personalized adjustment of the control strategy, and the trust estimation results are used as a supplementary adjustment factor to correct macro-decision-making such as route selection and to provide warnings for over-trusting drivers. Ultimately, a "personality adaptation + state response" fusion control mechanism is formed, which enables the autonomous driving system to ensure consistency in behavioral style while promptly responding to the collaborative risks brought about by trust fluctuations, thereby realizing a safer and more controllable human-machine co-driving control strategy.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for controlling an autonomous vehicle based on driving style and dynamic trust evaluation, the method comprising the following steps:

[0008] Step 1. Identifying the driver's driving style based on driving data during vehicle driving;

[0009] Step 2. Based on the identified driver's driving style, the FiLM-Gate style modulation mechanism is introduced to optimize the LSTM neural network. The speed and acceleration at the next moment are predicted based on the optimized LSTM neural network.

[0010] Step 3. Dynamically adjust the vehicle's PID control parameters based on the current driving style and predicted information, and generate execution instructions;

[0011] Step 4. Collect the driver's response characteristics to the autonomous driving behavior, and evaluate the driver's trust level in the autonomous vehicle in real time based on a partially observable Markov decision process. Use the trust evaluation results as a supplementary adjustment factor to make the optimal action decision. The style-trust dual-channel fusion control vehicle is controlled to achieve a joint closed-loop control of "personalization + trust perception + safety optimization".

[0012] As a preferred embodiment of the present invention, when identifying the driving style, a comprehensive driving style score DS is first calculated, and then the comprehensive driving style score DS is normalized and compared with a set threshold to determine the driver's driving style; the comprehensive driving style score DS is calculated based on the driving volatility DV and the following stability index FTI.

[0013] As a preferred embodiment of the present invention, step 2 specifically includes the following steps:

[0014] Step 2.1. Map the driving style label to a sparse vector representation through a one-hot encoding vector;

[0015] Step 2.2. Through the embedding matrix Convert sparse vectors into multi-dimensional continuous features, i.e., multi-dimensional style condition vectors;

[0016] Step 2.3. Process the multidimensional style condition vector obtained in step 2.2 using a two-layer perceptron to obtain the style hidden features.

[0017] Step 2.4. Calculate the scaling vector using style latent features , translation vector ;

[0018] Step 2.5. Input the original features of each vehicle into a single-layer LSTM neural network, using the scaled vector obtained in step 2.4 , translation vector Memory vector of the original candidate Perform style modulation to obtain the modulated candidate memory vector , the modulated candidate memory vector Enter the LSTM standard state update and output the new hidden state of time step t; then perform residual connection, layer normalization, shared full connection and nonlinear mapping in sequence to extract common features ;Finally, the predicted velocity and acceleration are output through a double-headed linear transformation.

[0019] As a preferred embodiment of the present invention, in step 3, the PID basic parameters are set to , extremely conservative PID control parameters The expression is:

[0020] ;

[0021] Conservative PID control parameters The expression is:

[0022] ;

[0023] Normal PID control parameters The expression is:

[0024] ;

[0025] Aggressive PID control parameters The expression is:

[0026] ;

[0027] Extremely aggressive PID control parameters The expression is:

[0028] ;

[0029] in, They are the style adjustment coefficients of extremely conservative, conservative, radical, and extremely radical, and their values ​​usually range from 0 to 1. Represents the acceleration at time t obtained in real time, is the predicted acceleration value, is the proportional coefficient, is the integral coefficient, is the differential coefficient.

[0030] As a preferred embodiment of the present invention, the partially observable Markov decision process POMDP in step 4 is in the following seven-tuple form:

[0031] ;

[0032] in, Represents a hidden trust state set; represents the set of actions taken by the autonomous vehicle; is the set of observable driver intervention behaviors; 、 、 are state transition probability, observation probability and immediate reward function respectively; is the discount factor;

[0033] Performing actions in an autonomous vehicle When calculating the driver's trust state transition probability, the style vector and environmental features are embedded; the observation probability is used to express the probability of a driver producing a specific observation behavior under specific trust state, action, driving style and environmental conditions;

[0034] The immediate reward function comprehensively considers efficiency, comfort and trust promotion, and introduces a tracking error penalty term based on the PID controller. Next action Instant rewards Defined as:

[0035] ;

[0036] in, , represents the weight coefficient, For vehicles in state Next action The efficiency improvement effect, The comfort improvement effect of the corresponding action is represented. Represents the positive promotion effect of the action on the driver's trust level; is the target acceleration tracking cost function.

[0037] As a further preferred embodiment of the present invention, the expression of the comprehensive driving style score DS is: ;in, and is a weight coefficient; the driving volatility DV is obtained by weighted summation of speed volatility, acceleration volatility, and azimuth volatility.

[0038] As a further preferred embodiment of the present invention, the modulated candidate memory vector The expression is: ;in, represents element-wise multiplication; the candidate memory vector after modulation Enter the LSTM standard state update, the expression is:

[0039] ;

[0040] ;

[0041] in, is the memory vector at time step t, is the new hidden state at time step t, is the time step The output gate.

[0042] As a further preferred embodiment of the present invention, the expression for generating the execution instruction in step 3 is:

[0043] ;

[0044] in, Represents the control amount of the throttle or brake, the acceleration error at time t , Represents the PID control parameters at time t.

[0045] As a preferred embodiment of the present invention, in step 4, the confidence state is used to represent the probability distribution of the driver's true trust state, and the confidence state is updated based on the Bayesian reasoning process, that is, each moment is updated based on the confidence state of the previous moment and the current observed behavior.

[0046] As a preferred embodiment of the present invention, the expression of the target acceleration tracking cost function in step 4 is:

[0047] ;

[0048] in, represents the PID control parameters at time t, is the integration time window, 、 are the acceleration errors at time t and time t-1 respectively.

[0049] Advantages and beneficial effects of the present invention:

[0050] (1) During vehicle operation, the present invention identifies the driving style based on the driver's real-time operating behavior, integrates the driving style identification results through the long short-term memory network (LSTM) structure, and generates or updates personalized control parameters that are more in line with the driver's expectations in real time; at the same time, the system continuously collects the driver's response characteristics to the autonomous driving behavior, and dynamically estimates the driver's trust status in the autonomous driving system based on the partially observable Markov decision process (POMDP) ​​modeling framework; the above-mentioned driving style identification results are used as the basis for personalized adjustment of the control strategy, and the trust estimation results are used as a supplementary adjustment factor for macro-decision corrections such as path selection and for warnings for drivers with excessive trust; ultimately, a "personality adaptation + state response" fusion control mechanism is formed, which enables the autonomous driving system to ensure the consistency of behavioral style while promptly responding to the collaborative risks brought about by trust fluctuations, thereby achieving a safer and more controllable human-machine co-driving control strategy.

[0051] (2) The present invention integrates two key elements: personalized recognition of driving style and dynamic evaluation of driver trust. By deeply perceiving the individual characteristics and changes in the driver's psychological state, it realizes the refinement and adaptive optimization of the control strategy of the autonomous driving vehicle, making the control strategy of the autonomous driving vehicle more accurate, efficient and intelligent. Compared with the existing control strategy, this method can effectively improve the overall perception ability of the autonomous driving control system of the driver's behavior and psychological state, avoid control conflicts, enhance the acceptability of the system, and significantly improve the safety level of human-machine co-driving.

[0052] (3) The present invention realizes real-time dynamic evaluation of the driver's trust status. A core challenge facing the widespread application of current autonomous driving technology is the user trust issue, that is, the degree of driver recognition and dependence on the decisions of autonomous vehicles. The driver's trust level is dynamic and is affected by the continuous interaction between the driving system behavior and the driving environment. Therefore, the present invention uses a partially observable Markov decision process (POMDP) ​​to construct a dynamic evaluation model of driver trust. That is, by dynamically capturing the implicit state changes of the driver's trust level, the system can judge and accurately adjust the degree of automated intervention in real time, actively adapt to the driver's psychological expectations and behavioral characteristics, thereby continuously consolidating and enhancing the driver's sense of trust, reducing safety risks caused by excessive or insufficient intervention, and comprehensively improving the reliability and user acceptance of autonomous driving vehicles.

[0053] (4) The present invention achieves personalized customization of the autonomous driving system and deeply optimizes the driving experience. Driving style is the driving preference and habit formed by drivers over a long period of time, and there are significant differences between different individuals. Traditional autonomous driving systems generally adopt standardized or universal decision logic, which makes it difficult to take into account the personalized driving demands of different drivers, resulting in differences in driving experience and subjective comfort, and even causing frequent driver intervention, which undermines system stability. To this end, the present invention proposes a refined driving style classification system based on driving volatility (DV) and following stability index (FTI), introduces the FiLM-Gate style modulation mechanism to optimize the LSTM neural network, and specifically predicts the driver's personalized behavioral needs and preferences. It dynamically optimizes the control parameters of the autonomous driving vehicle (such as PID control coefficients) in real time, so that the vehicle control strategy accurately matches the driver's individual characteristics and driving style requirements. This personalized control scheme can not only significantly improve the driver's comfort, pleasure and sense of security, reduce the subjective discomfort and operational interference caused by driving style differences, but also improve the driver's overall satisfaction and trust in autonomous driving technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of the autonomous driving vehicle control method based on driving style and dynamic trust evaluation of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application is described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0056] like Figure 1 As shown, this embodiment provides an autonomous driving vehicle control method based on driving style and dynamic trust evaluation, the method comprising the following steps:

[0057] Step 1. Identify the driver's driving style based on the driving data during vehicle driving:

[0058] Driving style refers to the driver's habits and attitude in performing driving tasks, which is mainly reflected in the driver's behavior and decision-making. The more aggressive the driver is, the more likely he is to engage in frequent dangerous driving behaviors.

[0059] In this embodiment, driving volatility (DV) is introduced. It is defined as a statistical measure of the changes in speed, acceleration, and azimuth during driving behavior. It reflects the instability and unpredictability of driving behavior and is used to improve the accuracy and detail of driving style classification.

[0060] Specifically, driving volatility (DV) is related to speed volatility, acceleration volatility, and azimuth volatility, and is calculated by the following steps:

[0061] 1) Calculate velocity volatility : , ; where N1 is the number of velocity data points, is the first-order difference of velocity, that is, the rate of change of velocity at the i-th time point, is the average value of the rate of change of velocity.

[0062] 2) Calculate the acceleration volatility : ; where N2 is the number of acceleration data points, is the acceleration at the i-th time point.

[0063] 3) Calculate the azimuthal volatility: ; where N3 is the number of azimuth data points, is the azimuth at the i-th time point.

[0064] 4) Calculate driving volatility: ;in, and is the weight coefficient, which can be adjusted according to the actual application scenario to reflect the relative importance of speed, acceleration and azimuth angle changes to driving volatility.

[0065] This embodiment introduces the Following Stability Index (FTI), which is defined as a metric for following vehicle behavior and considers two key factors: following distance and following speed. Following distance refers to the distance between the driver's vehicle and the vehicle ahead, typically measured in meters (or vehicle length), and can be measured using sensors such as radar, laser, or cameras. Following speed refers to the speed of the driver's vehicle relative to the vehicle ahead, typically measured in kilometers per hour. Changes in following speed can reflect the driver's adjustment to the speed of the vehicle ahead. Multiplying following distance stability and following speed stability can comprehensively reflect the driver's stability of following distance and following speed during the following process. If the driver can maintain both a stable following distance and a stable following speed during the following process, the Following Stability Index will be higher. The expression for the Following Stability Index (FTI) is:

[0066] ;

[0067] in: is the average following distance, is the standard deviation of following distances, is the average following speed, is the standard deviation of following vehicle speeds.

[0068] In this embodiment, the comprehensive driving style score is calculated based on the driving volatility and the following vehicle stability index. , the expression is: ;in, and is the weight coefficient.

[0069] The calculated DS is then normalized and compared with the set threshold to categorize the driver's driving style into the following categories:

[0070] If 0≤DS<15%, it is extremely conservative, with a fixed value of 1. The driving behavior is cautious and stable, tending to focus on comfort and safety.

[0071] If 15%≤DS<35%, it is conservative, the fixed value is 2, the driving behavior is relatively stable, and violent driving actions are avoided;

[0072] If 35%≤DS<65%, it is normal, the set value is 3, and the driving behavior is moderate, achieving a balance between efficiency and comfort;

[0073] If 65%≤DS<85%, it is aggressive, with a set value of 4, and the driving action is more decisive, with more speed or acceleration changes;

[0074] If 85%≤DS≤100%, it is extremely aggressive: the score range is fixed at 5, the driving style changes dramatically, and quick response is sought.

[0075] Step 2. Based on the identified driver's driving style, the FiLM-Gate style modulation mechanism is introduced to optimize the LSTM neural network. The speed and acceleration at the next moment are predicted based on the optimized LSTM neural network:

[0076] In this embodiment, the FiLM-Gate style modulation mechanism is introduced into the long short-term memory neural network (LSTM) to embed the driving style and record the real-time style label. ; Through One-hot encoding vector, Map to Sparse vectors of , that is, constructing standard basis vectors ,but Since one-hot is very sparse (99% are 0), it is difficult for neural networks to learn; the embedding matrix Compressing style information into continuous dimensions is more suitable for gradient optimization. ; For all windows of the same DS, take the mean value of 7 indicators respectively, , and get a 5×7 numerical matrix ;in, to Represents the mean of the above seven indicators when the style labels are 1 to 5. Due to the large differences in driving volatility and distance dimensions, the min-max method is used to scale to [0, 1] to ensure that the gradients in FiLM-MLP are of the same magnitude and the values ​​are monotonic. ;in, represents the normalized mean of the i-th driving style (DS=i) on the j-th driving index, Represented by a 5×7 driving style prototype matrix, the results Let the embedding dimension ,direct , you can complete the embedding table of "row = style prototype, column = seven indicators"; finally, we get ; Represents the generation of d emb dimensional style condition vector, one-hot vector d t Converting to 7-dimensional continuous features, the essence is to select the corresponding DS in M t The superscript ⊤ represents the transpose symbol.

[0077] In order to inject driving style information into the LSTM neural network, this embodiment introduces a set of two-layer perceptrons (MLP). Input the hidden layer with ReLU activation to obtain the style hidden features , the expression is:

[0078] ;

[0079] in, and , are the weight and bias of the first layer respectively, is the number of hidden units, and ReLU activation ensures feature sparsity and improves expressiveness.

[0080] Hiding features using style Calculate the scaling vector , the expression is:

[0081] ;

[0082] in, , to generate the scaling vector The weight of , is the bias of the scaling vector, is the LSTM hidden layer dimension.

[0083] Similarly, the translation vector is obtained by another linear projection , the expression is:

[0084] ;

[0085] in, , to generate the translation vector The weight of , is the offset of the translation vector.

[0086] When style modulation is not considered, the existing LSTM Will calculate the input gate , Forget Gate , output gate and the original candidate memory vector , the expression is:

[0087] ;

[0088] in, is the Sigmoid activation function, The input weights of the LSTM input gate, forget gate, output gate, and original candidate memory vector; is the recurrent weight of the LSTM gate, is the bias term of LSTM gating, is the input of LSTM, which contains the original features of the vehicle at each moment. , is the hidden state of the previous time step (i.e., time step t−1).

[0089] The FiLM-Gate style modulation mechanism is to use external conditions to perform element-by-element linear transformation on the intermediate features while keeping the network structure unchanged. Specifically for the LSTM neural network, in this embodiment, the original candidate memory vector Perform style modulation to obtain the modulated candidate memory vector , the expression is: ;in, Represents element-wise multiplication, scaling vector It is equivalent to applying a learnable scaling weight to the candidate memory component to adjust the sensitivity of each hidden unit to new information; the translation vector It provides a learnable bias term to determine the baseline offset written to the candidate state under different style conditions; both come from style embedding, so each memory update can adaptively reflect the differentiated characteristics of driving style, thereby achieving refined encoding of driving preferences at the internal representation level.

[0090] Modulated candidate memory vector Enter the LSTM standard state update, the expression is:

[0091] ;

[0092] ;

[0093] in, is the memory vector at time step t, is the new hidden state at time step t; thus, the same input sequence has different Under these conditions, Different stylized memory evolution and hidden states are generated, which in turn affect the final velocity and acceleration predictions.

[0094] Specifically, in this embodiment, represents the output hidden state vector of the Lth layer of LSTM (a single-layer LSTM is used in this embodiment, L=1), and is expressed as:

[0095] ;

[0096] in, is the memory vector of the L-th layer time step t, is the output gate activation vector of the Lth layer at time step t;

[0097] right First, make a residual connection and perform layer normalization to enhance training stability and gradient flow:

[0098] ;

[0099] ;

[0100] in, , is the residual mapping matrix, used to learn the linear transformation of the original hidden state; Represents the layer normalization operation, which performs normalization on each element of the vector. ,in and They are The mean and variance of is a small constant that ensures numerical stability. is the intermediate vector after adding the residual, is the final output after normalization and residual.

[0101] Then, a layer of shared full connection and nonlinear mapping is used to extract common features , the expression is:

[0102] ;

[0103] in, Select ReLU activation function; , is the weight matrix of the shared feature extraction layer, Hidden dimension for regression; , is the bias vector of the shared feature extraction layer.

[0104] Finally, the double-head linear mapping is used to predict velocity and acceleration respectively:

[0105]

[0106] in, , are the weight vectors of velocity and acceleration regression heads respectively; , are the scalar bias terms of the velocity and acceleration regression heads respectively; , is the estimated value of the longitudinal velocity at the next moment; , is the estimated value of the longitudinal acceleration at the next moment.

[0107] In this embodiment, the LSTM neural network is trained by using historical data and referring to the existing method. After the training, the optimal weight set obtained by training is (including embedding matrix , FiLM-MLP weights, LSTM parameters and output layer weights) are packaged into the vehicle-side executable model FiLM-Gate-LSTM ( ); deployed in the autonomous vehicle decision system z2, it completes a forward inference at a cycle of 10 Hz (every 100 ms) during operation and outputs .

[0108] Step 3. Dynamically adjust the vehicle's PID control parameters based on the current driving style and predicted information, and generate execution instructions:

[0109] In this embodiment, the predicted value As the target acceleration, read the current driving style tag , for the above five driving styles, the present invention proposes personalized PID control parameters (proportional coefficient , integral coefficient , differential coefficient ) Dynamically adjust strategies.

[0110] Specifically, let the PID basic parameters be , the specific control parameter adjustment formula for each style is as follows:

[0111] Extremely conservative: To ensure comfort and stability, reduce the sensitivity of the response to avoid frequent and large control actions, and define extremely conservative PID control parameters The expression is:

[0112] ;

[0113] Specifically, let the PID basic parameters be , the specific control parameter adjustment formula for each style is as follows:

[0114] Extremely conservative: To ensure comfort and stability, reduce the sensitivity of the response to avoid frequent and large control actions, and define extremely conservative PID control parameters The expression is:

[0115] ;

[0116] Normal type: Use the baseline control parameters to strike a balance between comfort and responsiveness, and define the normal PID control parameters The expression is:

[0117] ;

[0118] Aggressive: Enhance response speed, improve vehicle agility, and define aggressive PID control parameters The expression is:

[0119] ;

[0120] Extremely aggressive: further improves the response sensitivity of the control system, allows larger control fluctuations to achieve rapid adaptation, and defines extremely aggressive PID control parameters The expression is:

[0121] ;

[0122] In the above formula, They are the style adjustment coefficients of extremely conservative, conservative, radical, and extremely radical, and their values ​​usually range from 0 to 1. Represents the acceleration at time t obtained in real time. This method ensures that the vehicle control parameters can be dynamically adjusted according to the driver's real-time driving style, realizing personalized vehicle control strategies for different styles, and effectively improving the adaptability and safety performance of autonomous vehicles.

[0123] In this embodiment, the above style adjustment coefficient can be output through the LSTM model, that is, a new output branch is added to the LSTM model to output the style adjustment coefficient in real time. The input of this branch is the style latent feature , use sigmoid activation function, ensure the coefficient range is [0, 1], and output style adjustment coefficient .

[0124] In this embodiment, the acceleration error is calculated , and generate execution instructions through the PID formula:

[0125] ;

[0126] in, Represents the control amount of the throttle or brake to ensure the actual acceleration of the vehicle Tracking Target .

[0127] Through this refined style classification control strategy, the vehicle control system can more accurately match the driver's desired behavior in various driving scenarios, comprehensively improving the driving experience and safety performance. Firstly, by introducing FiLM-Gate modulation, real-time style labels are injected into the LSTM, ensuring that the predicted target aligns with the driver's preferences. Secondly, the same style label is used to fine-grainedly scale the PID coefficients, ensuring that the vehicle exhibits appropriate acceleration and deceleration feel and comfort when achieving the same target.

[0128] Step 4: Evaluate the driver's trust in the autonomous vehicle in real time, use the trust evaluation result as a supplementary adjustment factor, and control the vehicle through style-trust dual-channel fusion:

[0129] In this embodiment, a real-time trust assessment model for human-computer interaction is established based on a partially observable Markov decision process (POMDP). On the basis of a control strategy generated by considering driving style, the trust assessment result is introduced as a supplementary adjustment factor, realizing a style-trust dual-channel fusion control mechanism.

[0130] In human-machine collaborative autonomous driving scenarios, a driver's level of trust in the autonomous vehicle directly impacts the effectiveness and safety of their driving behavior. Assuming a driver's style is identified as "conservative," the LSTM module will configure conservative parameters such as low acceleration and low speed. However, if the driver frequently intervenes or hesitates, the system will estimate a decrease in trust in real time. In this case, even if the system is configured in "conservative mode," further measures should be taken: slowing down lane change responses, increasing the intensity of human-machine interface (HMI) prompts, or even pausing the current action to prompt the driver to intervene. However, this trust cannot usually be directly and accurately measured, but it exhibits significant partial observability. Therefore, this paper introduces the Partially Observable Markov Decision Process (POMDP) ​​as the theoretical basis for the trust model. The POMDP is applicable to environments with uncertainty, partially observable states, and random behavior, and can effectively describe the dynamic evolution of trust during the interaction between the driver and the autonomous driving system. The Partially Observable Markov Decision Process (POMDP) ​​has significant advantages in characterizing the dynamic evolution of a driver's trust in the autonomous vehicle. The POMDP model is defined as the following seven-tuple form:

[0131] ;

[0132] in, Represents a hidden trust state set; represents the set of actions taken by the autonomous vehicle; is the set of observable driver intervention behaviors; 、 、 are state transition probability, observation probability and immediate reward function respectively; is the discount factor.

[0133] Specifically, the driver's actual trust in the autonomous vehicle is an implicit random variable that cannot be directly observed. In this embodiment, the hidden trust state set is defined as: , indicating the driver’s insufficient trust, moderate trust, and excessive trust, respectively.

[0134] The action set is defined as: , respectively representing continuing the current path, switching to a conservative path (avoiding the main road), switching to an efficient path (allowing more aggressive lane changes), requesting the driver to take over, triggering the current decision explanation and attention deviation warning, using one-hot encoding, the selected action is encoded as 1, and the rest of the actions are 0; the vector after one-hot encoding at time t is defined as the decision action vector .

[0135] The set of observed driver intervention behaviors is defined as: , which represent the proportion of time spent looking at the road, the intensity of glances, and the frequency of intervention, respectively. They are collected in real time by the in-car sensor. The three observations are in a fixed-length sliding time window. Statistics within 5 seconds: the proportion of time spent looking at the road at time t , scanning intensity and intervention frequency The expressions are: ;in, for The total time the driver keeps his eyes on the "driving area ahead"; ;in, Counts the gaze area switches detected by the eye tracker. Experience limit 20; ;in, The number of intervention actions such as steering wheel takeover and brake pedal intervention, The experience limit is 3.

[0136] In this embodiment, the driving style embedding vector uses the style vector constructed in step 2 ;Context vector provided by the environment perception module = , are divided into three levels: 1, 2, and 3. Yes The feature representation after linear transformation and nonlinear activation is , is the trainable weight matrix, is the bias vector, is a nonlinear activation function. is the environment context vector The output dimension of .

[0137] Performing actions in an autonomous vehicle Driver trust status Transition to the next state The expression of the state transition probability is:

[0138] = ;

[0139] in, , The basic state transition probability parameter represents the default state transition tendency when there is no style and environment influence. It is a learnable parameter. and Used to dynamically adjust the state transition probability. If the driving style is more aggressive (DS=4,5), the probability of trust drop increases under aggressive control behaviors (acceleration, frequent lane changes). If the environment is complex (dense traffic, rainy days), the probability of trust drop increases under the same conditions. is the adjustment weight vector of driving style on state transition, is the adjustment weight vector of the environmental context on the state transition.

[0140] The observation probability represents the probability of a driver producing a specific observation behavior under specific trust state, action, driving style and environmental conditions, and is expressed as:

[0141] = ;

[0142] in, , is the basic observation probability parameter, which indicates the default observation tendency when there are no style and environmental factors. It is a learnable parameter. The weight vector for adjusting the driving style to observation probability, The weight vector for adjusting the observation probability for the environmental context, It is a continuous observation space used for integral normalization. Since the true trust state cannot be directly measured, this paper uses the belief state to represent the probability distribution of the driver's true trust state, and defines the belief state as:

[0143] ;

[0144] in, is all driver intervention behaviors collected up to time t; is the vehicle control action executed by the system from step 1 to t−1; is the belief state space, The confidence state is updated based on the Bayesian reasoning process. Each moment is updated based on the confidence state of the previous moment and the current observation behavior. Execute and get new observations After that, the confidence status is updated to:

[0145] = ;

[0146] Among them, the denominator represents the marginal probability of observing the current observation in all possible states, which is a normalization constant. represents all possible trust states at time t+1, Represents all possible trust states at time t.

[0147] The immediate reward function comprehensively considers efficiency, comfort, and trust promotion, and introduces a tracking error penalty term based on a PID controller, which is defined as:

[0148] ;

[0149] in, , Represents the weight coefficient, which can be adaptively adjusted. That is, by performing softmax normalization on the instantaneous reward signals of the four dimensions, the relative importance of each dimension in the overall reward at the next moment is dynamically determined;

[0150] For vehicles in state Next action The efficiency improvement effect, ,in, Refers to the system in its current state Select Action After that, the remaining time required to reach the next track control point or complete the current obstacle avoidance task is predicted; The maximum remaining travel time threshold allowed in similar scenarios is used to Normalized, its size depends on the road speed limit, traffic density and regulatory constraints, with typical values ​​of 20–40 s. Indicates execution of an action The additional energy consumption in the last 5 s window relative to the energy-saving baseline condition, Indicates the upper limit of energy consumption increment allowed under the same road conditions and vehicle specifications; when a vehicle cuts in from the side, the definition is the net distance difference between the vehicle and the inserted vehicle at the moment of longitudinal coordinate alignment, is a coefficient that can be adjusted according to actual conditions; Sigmoid is used to convert Mapped to [0,1].

[0151] represents the comfort improvement effect of the action (based on vehicle control smoothness evaluation), , ,in 、 is the root mean square value of the longitudinal and lateral acceleration, which measures the dynamic intensity of the vehicle during this period. is the root mean square value of the first derivative of longitudinal acceleration (jerk), which is used to evaluate the smoothness of acceleration change. 、 They are the longitudinal and lateral comfort acceleration thresholds, respectively. Experts recommend setting the threshold to 3 m / s². Exceeding the threshold will result in cumulative penalties in the comfort index. is the comfort threshold of jerk, usually 1 m / s³, is a learnable parameter, is the weight coefficient, which can be adjusted according to actual conditions.

[0152] Represents the positive promotion effect of the action on the driver's trust level; each discrete trust state Mapping to intervals Scale on (0 to 1, the degree of trust gradually increases), commonly used equal interval or empirical scale, ; Let the current confidence distribution be , then the weighted expectation gets a continuous scale Both excessive and insufficient trust can be dangerous. For the optimal point, define the normalized distance It means complete moderation; indicates extreme imbalance (extremely low or excessive trust); , is a learnable parameter, is a coefficient that can be adjusted according to actual conditions.

[0153] is the target acceleration tracking cost function, is the set of three PID parameters at time t, when When it increases, the trust state tends to shift toward “low trust,” triggering more conservative or interpretive actions. , ;in, represents the PID control parameters at time t, is the integration time window, usually the sampling period of the controller.

[0154] The goal of the model is to find an optimal strategy , with the goal of maximizing the expected cumulative discounted reward, the specific optimization objective function is defined as:

[0155] ;

[0156] in, represents the optimal expected utility value under the current confidence state, is the discount factor used to discount the impact of future rewards, Represents the probability of observation at the next moment under the current confidence state and action:

[0157] = , is a new belief vector obtained by a Bayesian update.

[0158] The optimal value function is solved iteratively until convergence:

[0159] ;

[0160] in, is the ∞-norm difference between the new and old value functions, For all confidence distributions Take the maximum absolute difference, is the convergence threshold, set . Update repeatedly until the convergence condition is met. The maximum expected cumulative reward (optimal expected utility value) that can be obtained by following the optimal strategy under the current confidence state can be obtained. is the decision action that should be performed under the current confidence state.

[0161] In this embodiment, the style vector output in step 2 is Together with the vehicle control parameters (PID coefficients) and control behavior data output from step 3, which serve as important regulating variables for the state transition probability and observation probability of this step, the driver's confidence state estimation and action decision output from this step further supplement and adjust the system decision to achieve the joint closed-loop control of "personalization + trust perception + safety optimization". These two steps jointly achieve the dual optimization of "behavior matching driver style" and "dynamic trust perception and adaptation" of autonomous driving vehicles, improving user experience and safety performance.

[0162] In this embodiment, the "behavior matching" and "trust perception" channels are tightly coupled. Channel 1, centered around an embedded-style FiLM-Gate LSTM network, predicts the optimal target acceleration and PID control parameters in real time, ensuring that the vehicle's motion style closely matches the driver's preferences. Channel 2, based on POMDP confidence reasoning, updates the trust distribution based on observed intervention behaviors, driving optimal control decisions and promptly strengthening safety constraints. For example, if the driver's trust in the system is assessed to be insufficient, the system avoids dangerous paths and switches to conservative routes (avoiding main roads), preparing for the driver's potential to take over the system at any time. If the driver's trust in the system is assessed to be excessive, the system warns that the driver's attention has completely shifted, requiring timely warnings of attention shifts. The two outputs collaboratively optimize and adjust vehicle control commands, enabling not only precise matching of driving style but also adaptive perception and maintenance of driver trust, thereby achieving a closed-loop collaboration of personalization, trust perception, and safety optimization.

[0163] This method can be integrated into L3 and above autonomous driving systems, and is suitable for real-time applications on urban roads and highways. It can effectively enhance driver trust and user satisfaction, reduce unnecessary takeover interventions, improve driving comfort and safety, and provide a feasible technical path and engineering implementation plan for the personalized, humanized and large-scale promotion and application of autonomous driving vehicles.

[0164] The present invention also provides an electronic device comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned autonomous driving vehicle control method based on driving style and dynamic trust assessment.

[0165] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned autonomous driving vehicle control method based on driving style and dynamic trust assessment.

[0166] Those skilled in the art will appreciate that all or part of the functions of the various methods / modules in the above embodiments may be implemented via hardware or via computer programs. When all or part of the functions in the above embodiments are implemented via computer programs, the program may be stored in a computer-readable storage medium, which may include a read-only memory, random access memory, a magnetic disk, an optical disk, a hard disk, etc., and the program is executed by a computer to implement the above functions. For example, the program may be stored in a memory of a device, and when the program in the memory is executed by a processor, all or part of the above functions may be implemented.

[0167] In addition, when all or part of the functions in the above-mentioned embodiments are implemented by means of a computer program, the program can also be stored in a storage medium such as a server, another computer, a disk, an optical disk, a flash drive or a mobile hard disk, and saved to the memory of a local device by downloading or copying, or the system of the local device is updated. When the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

[0168] The above description of the present invention using specific examples is intended only to facilitate understanding of the present invention and is not intended to limit the present invention. A person skilled in the art of the present invention may make several simple deductions, modifications, or substitutions based on the principles of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for controlling an autonomous vehicle based on driving style and dynamic trust evaluation, characterized in that: The method comprises the following steps: Step 1. Identifying the driver's driving style based on driving data during vehicle driving; Step 2. Based on the identified driver's driving style, the FiLM-Gate style modulation mechanism is introduced to optimize the LSTM neural network. The speed and acceleration at the next moment are predicted based on the optimized LSTM neural network. Step 3. Dynamically adjust the vehicle's PID control parameters based on the current driving style and predicted information, and generate execution instructions; Step 4. Collect the driver's response characteristics to the autonomous driving behavior, and evaluate the driver's trust level in the autonomous vehicle in real time based on a partially observable Markov decision process. Use the trust evaluation results as a supplementary adjustment factor to make the optimal action decision. Control the vehicle through the style-trust dual-channel fusion to achieve a joint closed-loop control of "personalization + trust perception + safety optimization." 2. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 1, characterized in that: When identifying the driving style, a comprehensive driving style score DS is first calculated, and then the comprehensive driving style score DS is normalized and compared with a set threshold to determine the driver's driving style; the comprehensive driving style score DS is calculated based on the driving volatility DV and the following stability index FTI.

3. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.

1. Map the driving style label to a sparse vector representation through a one-hot encoding vector; Step 2.

2. Through the embedding matrix Convert sparse vectors into multi-dimensional continuous features, i.e., multi-dimensional style condition vectors; Step 2.

3. Process the multidimensional style condition vector obtained in step 2.2 using a two-layer perceptron to obtain the style hidden features. Step 2.

4. Calculate the scaling vector using style latent features , translation vector ; Step 2.

5. Input the original features of each vehicle into a single-layer LSTM neural network, using the scaled vector obtained in step 2.4 , translation vector Memory vector of the original candidate Perform style modulation to obtain the modulated candidate memory vector , the modulated candidate memory vector Enter the LSTM standard state update and output the new hidden state of time step t; then perform residual connection, layer normalization, shared full connection and nonlinear mapping in sequence to extract common features ;Finally, the predicted velocity and acceleration are output through a double-headed linear transformation.

4. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 1, characterized in that: In step 3, the basic PID parameters are set to , extremely conservative PID control parameters The expression is: ； Conservative PID control parameters The expression is: ； Normal PID control parameters The expression is: ； Aggressive PID control parameters The expression is: ； Extremely aggressive PID control parameters The expression is: ； in, They are the style adjustment coefficients of extremely conservative, conservative, radical, and extremely radical, with a value range between 0 and 1. Represents the acceleration at time t obtained in real time, is the predicted acceleration value, is the proportional coefficient, is the integral coefficient, is the differential coefficient.

5. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 1, characterized in that: The partially observable Markov decision process POMDP in step 4 is in the following seven-tuple form: ； in, Represents a hidden trust state set; represents the set of actions taken by the autonomous vehicle; is the set of observable driver intervention behaviors; 、 、 are state transition probability, observation probability and immediate reward function respectively; is the discount factor; Performing actions in an autonomous vehicle When calculating the driver's trust state transition probability, the style vector and environmental features are embedded; the observation probability is used to represent the probability of the driver generating a specific observation behavior under specific trust state, action, driving style and environmental conditions; The immediate reward function comprehensively considers efficiency, comfort and trust promotion, and introduces a tracking error penalty term based on the PID controller. Next action Instant rewards Defined as: ； in, , represents the weight coefficient, For vehicles in state Next action The efficiency improvement effect, The comfort improvement effect of the corresponding action is represented. Represents the positive promotion effect of the action on the driver's trust level; is the target acceleration tracking cost function.

6. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 2, characterized in that: The expression of comprehensive driving style score DS is: ;in, and is a weight coefficient; the driving volatility DV is obtained by weighted summation of speed volatility, acceleration volatility, and azimuth volatility.

7. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 3, characterized in that: Modulated candidate memory vector The expression is: ;in, represents element-wise multiplication; the candidate memory vector after modulation Enter the LSTM standard state update, the expression is: ； ； in, is the memory vector at time step t, is the new hidden state at time step t, is the time step The output gate.

8. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 4, characterized in that: The expression for generating the execution instruction in step 3 is: ； in, Represents the control amount of the throttle or brake, the acceleration error at time t , Represents the PID control parameters at time t.

9. The method for controlling an autonomous vehicle based on driving style and dynamic trust evaluation according to claim 5, wherein: In step 4, the confidence state is used to represent the probability distribution of the driver's true trust state, and the confidence state is updated based on the Bayesian reasoning process, that is, each moment is updated based on the confidence state of the previous moment and the current observed behavior.

10. The autonomous driving vehicle control method based on driving style and dynamic trust evaluation according to claim 5, characterized in that: The expression of the target acceleration tracking cost function in step 4 is: ； in, represents the PID control parameters at time t, is the integration time window, 、 are the acceleration errors at time t and time t-1 respectively.

Images