A humanized following vehicle driving method based on stochastic model predictive control

Abstract

This invention discloses a human-like following-vehicle driving method based on stochastic model predictive control, relating to the field of environmental perception and decision-making technology for autonomous vehicles. The method includes: creating training and testing sets; fitting a following-vehicle model; constructing a driving style classifier; training and testing the driving style classifier to obtain a driving style recognition model; inputting real-time driving data into the driving style recognition model to adjust the driving mode; learning and optimizing Gaussian process hyperparameters based on the training and testing sets to obtain hyperparameter learning results; establishing a vehicle motion state predictor; establishing a state equation and optimizing the objective function; binding the state variables with the prediction results through probabilistic constraints to obtain control variables; and controlling the vehicle based on the control variables. This invention considers the following-vehicle driving characteristics of the driver and the uncertainty of the preceding vehicle's motion, achieving safe, stable, and comfortable following-vehicle driving to improve the human-likeness and acceptability of intelligent vehicles.

Description

Technical Field

[0001] This invention relates to the field of environmental perception and decision-making technology for autonomous vehicles, and more specifically, to a human-like following driving method based on stochastic model predictive control. Background Technology

[0002] The development of intelligent vehicles is a gradual process. According to the autonomous driving classification standards defined by the Society of Automotive Engineers (SAE), autonomous vehicles are divided into six levels, L0-L5. Based on current technological conditions, intelligent vehicles will remain at levels L2-L3 for a considerable period in the future, meaning that the driver and the vehicle will jointly complete the driving task (human-machine co-driving).

[0003] Currently, the initial design goals of mass-produced L2-L3 level autonomous driving systems are to reduce driver workload and prevent traffic accidents. In these systems, the driver still bears significant driving responsibilities, resulting in prolonged human-machine interaction during vehicle operation. However, most systems prioritize functional feasibility while neglecting the driver's individual needs. This leads to a rigid and monotonous operating process, resulting in low driver acceptance. A good following strategy requires automatically adjusting the driving mode based on different driving styles, ensuring driving safety—essentially, "adapting the car to the driver." In actual driving, the performance of the lead vehicle is also affected by the movement of the vehicle in front. Therefore, considering driving safety and comfort, how to account for driving style and the uncertainty of obstacle movement during following has become a research challenge in the field of advanced autonomous driving. Summary of the Invention

[0004] This invention provides an anthropomorphic following driving method based on stochastic model predictive control, in order to overcome at least one technical problem existing in the prior art.

[0005] This invention provides a human-like following driving method based on stochastic model predictive control, comprising:

[0006] Historical driving data under following conditions is collected, and the historical driving data is calibrated to obtain various calibration data and driver styles; the driver styles include at least cautious drivers, general drivers, and aggressive drivers;

[0007] The calibration data was used to create training and testing sets;

[0008] The following model is fitted using the training set and the test set to obtain a driving mode corresponding to the driver's style; the following model is d. des =τv e +d0, where ddes v represents the driver's desired distance from other vehicles. e The driving modes include at least a cautious driving mode, a normal driving mode, and an aggressive driving mode. The driving modes represent the longitudinal speed of the main vehicle, d0 represents the minimum safe distance, and τ represents the driver's desired headway.

[0009] Select driving style feature data from the historical driving data, and use the driving style feature data to construct a driving style classifier;

[0010] The driving style classifier is trained and tested using the training set and the test set to obtain a driving style recognition model;

[0011] Real-time driving data under following conditions is collected, and the real-time driving data is input into the driving style recognition model to obtain the driver style. The driving mode is then adjusted according to the driver style to obtain the final driving mode.

[0012] The mean function and covariance function are defined using a Gaussian process f(x) = GP(x(x), k(x, z)); where f(x) represents a Gaussian process, x and z represent the inputs, x(x) represents the mean function, and k(x, z) represents the covariance function.

[0013] Assume from x0 to x n At time t, the input is x = {x0, x1, ..., x2} n The vehicle's motion state is y = {y0, y1, ..., y}. n If the predicted input is x′ and the predicted motion state is y′, then y follows a multivariate Gaussian distribution, described as follows: Where k(x, x) is the covariance function with respect to the training set, k(x, x′) is the covariance function between the training and test sets, and k(x′, x′) is the covariance function with respect to the test set; N() represents a Gaussian distribution with a mean of 0. This represents environmental noise parameters, and I represents the identity matrix.

[0014] Calculate the marginal distribution of y′, and obtain the expression for the uncertainty of the prediction result as follows: Where, y′ mean This represents the prediction result, y′ cov This represents the covariance of the predicted values;

[0015] The hyperparameters of the Gaussian process are learned and optimized based on the training set and the test set. The hyperparameter learning results are obtained according to the log-likelihood function and the partial derivatives of each hyperparameter.

[0016] Based on the learning results, a vehicle motion state predictor is established to obtain a prediction result of the motion state of the vehicle in front. The prediction result of the motion state of the vehicle in front includes at least a prediction result of speed and a prediction result of acceleration.

[0017] Establish state equations and optimize objective functions to ensure that the following vehicle driving process meets the needs of comfort, safety, and driver requirements;

[0018] The prediction results are then constrained using probabilistic constraints to obtain control variables;

[0019] The vehicle is controlled based on the control variables.

[0020] Optionally, training and testing sets are created using the calibration data, specifically as follows:

[0021] The calibration data is divided into a training set and a test set according to a predetermined ratio.

[0022] Optionally, the following conditions include at least lane restrictions, distance restrictions, speed restrictions, and time restrictions;

[0023] Among them, lane restrictions include: when the main vehicle and the vehicle in front are traveling in the same lane, the following condition ends when another vehicle cuts into or out of the space between the two vehicles;

[0024] Distance restrictions include: the relative distance between the two vehicles is S, and 5m ≤ S ≤ 120m;

[0025] The speed limits include: for both vehicles, the speed is v, and 20km / h ≤ v ≤ 120km / h;

[0026] Time restrictions include: following the vehicle for more than 30 seconds.

[0027] Optionally, driving style feature data can be selected from the historical driving data, specifically:

[0028] Correlation analysis was performed on the historical driving data to obtain correlation values;

[0029] After removing data whose correlation values ​​are greater than the threshold, the driving style feature data is obtained as [v e a e [S, Δv, ξ], where v e Indicates the longitudinal speed of the main vehicle, a e ξ represents the longitudinal acceleration of the main vehicle, S represents the relative distance, Δv represents the relative velocity, and ξ represents the actual headway.

[0030] Optionally, This represents the variance of the signal controlling the output amplitude. The signal variance representing the control input amplitude; the hyperparameter is...

[0031] Optionally, the marginal distribution of y′ is: P(y′|x,x′,y)=N(y′) mean y′ cov ),in,

[0032] Optionally, the hyperparameter learning results are obtained based on the log-likelihood function and the partial derivatives of each hyperparameter, specifically:

[0033] Using the maximum likelihood method, through the formula Calculate the hyperparameters of the Gaussian process, where L(x, ψ) represents the log-likelihood function.

[0034] Optionally, the state equations are m(t+1)=Am(t)+Bu(t)+Gω(t), n(t)=Cm(t), where m represents the state variables, m=[Δd,Δv,a e ,j]′,Δd=Sd des Δd represents the distance error, Δv represents the relative speed, and a e Let j represent the longitudinal acceleration of the main vehicle, j represent the impact of the main vehicle, n represent the output variable, u represent the control variable, ω represent the disturbance caused by the motion state of the preceding vehicle, A is the system matrix, B is the control matrix, C is the output matrix, and G is the noise figure matrix.

[0035] The comfort expression is in, The weighting coefficient representing acceleration, γ j Weighting coefficients representing the degree of impact;

[0036] Driver requirements include driver dynamic characteristics, which are expressed as J. d =γ Δd Δd 2 +γ Δv Δv 2 , where γ Δd γ represents the weighting coefficient of the distance error. Δv Weighting coefficients representing relative velocity;

[0037] The expression for security is: Among them, t TTC Indicates the collision time, d safe Indicates the safe following distance, d s Indicates the maximum safe distance;

[0038] The objective function expression is optimized as follows: Where st represents a constraint, J total This represents the total cost function.

[0039] Optionally, the prediction results are constrained to bind the state variables using probabilistic constraints to obtain control variables, specifically:

[0040] By using probability constraints P(A) s m≤b s The state variables are bound together with ≥β, where A s Denotes the state constraint matrix, b s The constraint range is represented by β, and the risk probability factor is represented by β; the state variable expression is: Let represent the deterministic part of the system, and e(t) represent the perturbation part of the system;

[0041] The deterministic portion is controlled using feedback control gain, and the disturbance portion is compensated using a model predictive control algorithm to obtain the control variable. Where g represents the control input provided by the model predictive control algorithm, and K lqr This represents the feedback control gain that enables the system (A, B) to stabilize.

[0042] Optionally, disturbances caused by the motion of the vehicle in front. in, This indicates the determined portion of the disturbance. This represents the probability of the disturbance.

[0043] The innovative aspects of this invention include:

[0044] 1. In this embodiment, the following driving characteristics of the driver and the uncertainty of the movement of the vehicle in front are taken into account. The vehicle drive / braking system can be controlled according to the movement status information of the main vehicle and the vehicle in front to achieve safe, stable and comfortable following driving, thereby improving the anthropomorphism and acceptability of intelligent vehicles. This is one of the innovations of this embodiment.

[0045] 2. In this embodiment, the online driving style recognition model enables the following process to automatically match the driving mode according to the current driver style. In response to the impact of unbounded disturbances caused by the random movement of the preceding vehicle on the performance of the master vehicle, the Gaussian process is used to predict the uncertainty of the preceding vehicle's movement in real time. The prediction results are used to tighten the state variables in real time through probability constraints, so that the feasible domain of the controller is always kept within a reasonable range, thereby improving the tracking accuracy and comfort of the controller. This is one of the innovative points of this embodiment. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart of a following vehicle driving method provided in an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of a vehicle following operation provided in an embodiment of the present invention;

[0049] Figure 3 This is a flowchart illustrating the selection of driving style feature data provided in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0052] This invention discloses a human-like following driving method based on stochastic model predictive control. The following sections provide detailed descriptions.

[0053] Figure 1 This is a flowchart of a following vehicle driving method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a vehicle-following operation provided in an embodiment of the present invention. Please refer to it. Figure 1 and Figure 2 The anthropomorphic following driving method based on stochastic model predictive control provided in this embodiment includes:

[0054] Step 1: Collect historical driving data under following conditions, calibrate the historical driving data to obtain various calibration data and driver styles; driver styles include at least cautious drivers, general drivers, and aggressive drivers;

[0055] Step 2: Create training and test sets using the calibration data;

[0056] Step 3: Fit the following model using the training and test sets to obtain the driving mode corresponding to the driver's style; the following model is d. des =τv e +d0, where d des v represents the driver's desired distance from other vehicles. e The driving modes include at least a cautious driving mode, a normal driving mode, and an aggressive driving mode. d0 represents the minimum safe distance, and τ represents the driver's desired headway.

[0057] Step 4: Select driving style feature data from historical driving data and use the driving style feature data to build a driving style classifier;

[0058] Step 5: Train and test the driving style classifier using the training and test sets to obtain the driving style recognition model;

[0059] Step 6: Collect real-time driving data under following conditions, input the real-time driving data into the driving style recognition model to obtain the driver style, adjust the driving mode according to the driver style, and obtain the final driving mode.

[0060] Step 7: Define the mean function and covariance function using the Gaussian process f(x) = GP(x(x), k(x, z)); where f(x) represents the Gaussian process, x and z represent the inputs, x(x) represents the mean function, and k(x, z) represents the covariance function.

[0061] Step 8: Assume from x0 to x n At time t, the input is x = {x0, x1, ..., x2} n The vehicle's motion state is y = {y0, y1, ..., y}. n If the predicted input is x′ and the predicted motion state is y′, then y follows a multivariate Gaussian distribution, described as follows: Where k(x, x) is the covariance function with respect to the training set, k(x, x′) is the covariance function between the training and test sets, and k(x′, x′) is the covariance function with respect to the test set; N() represents a Gaussian distribution with a mean of 0. This represents environmental noise parameters, and I represents the identity matrix.

[0062] Step 9: Calculate the marginal distribution of y′ to obtain the uncertainty expression of the prediction result. Where, y′ mean This represents the prediction result, y′ cov This represents the covariance of the predicted values;

[0063] Step 10: Learn and optimize the hyperparameters of the Gaussian process based on the training set and the test set. Obtain the hyperparameter learning results based on the log-likelihood function and the partial derivatives of each hyperparameter.

[0064] Step 11: Based on the learning results, establish a vehicle motion state predictor to obtain the prediction results of the motion state of the vehicle in front. The prediction results of the motion state of the vehicle in front shall include at least the prediction results of speed and acceleration.

[0065] Step 12: Establish the state equation and optimize the objective function to ensure that the following vehicle process meets the needs of comfort, safety and driver requirements;

[0066] Step 13: Tighten the state variables by applying probabilistic constraints to the prediction results to obtain the control variables;

[0067] Step 14: Control the vehicle according to the control variables.

[0068] For details, please refer to Figure 1 and Figure 2 The anthropomorphic following-driving method based on stochastic model predictive control provided in this invention collects a large amount of natural driving information from drivers under different driving environments through information collection devices such as cameras and radar. This natural driving information includes vehicle-related information and environmental information. Driving environments include highways, intercity expressways, and national highways. Vehicle-related information includes information about the driver's own vehicle and the vehicle ahead. Specifically, the driver's own vehicle information includes vehicle speed, acceleration, steering wheel angle, yaw rate, and position information, while the vehicle ahead information includes relative speed and position information. Furthermore, the driver should be a normal driver with over 20,000 kilometers of driving experience, representing different occupations and age groups.

[0069] After obtaining the natural driving information, in this embodiment, historical driving data under following conditions is extracted from the natural driving information through step 1. The historical driving data is calibrated according to the characteristics of the driving data to obtain calibrated driving data. At the same time, the driver is divided into different styles, such as cautious driver, normal driver and aggressive driver.

[0070] After obtaining the calibration data, in step 2, the calibration data is divided into a training set and a test set according to a predetermined ratio. The training set includes the training sets for the online driving style recognizer and the vehicle state predictor, and the test set includes the test sets for the online driving style recognizer and the vehicle state predictor. The predetermined ratio can be, for example, 7:3, 8:2, 9:1, etc. Different ratios result in different amounts of data in the training set. When building the training model, the more data in the training set, the more thorough the training, and the higher the accuracy of the trained model, which is beneficial to improving the accuracy of the prediction results. However, at the same time, the more data in the training set, the greater the training complexity will be. Therefore, in practical use, the specific predetermined ratio can be selected as needed, and this application does not impose a specific limitation on it.

[0071] In step 3, the following model is fitted using the training and test sets, where the following model expression is d. des =τv e +d0, d des v represents the driver's desired distance from other vehicles. e Let d0 represent the longitudinal speed of the main vehicle, τ represent the minimum safe distance, and τ represent the driver's desired headway. Since the calibration data contains different types, three following models corresponding one-to-one with the calibration data can be obtained after fitting, each corresponding to one of the three driving modes: cautious driving mode, normal driving mode, and aggressive driving mode. Thus, in practical applications, by identifying the driver's style and adjusting the driving mode accordingly, the driving mode can be matched to the current driver's style.

[0072] Driver style needs to be identified using a driving style recognition model; therefore, a driving style recognition model needs to be constructed first. This invention first establishes a driving style classifier in step 4. To improve computational efficiency and reduce training model costs, the amount of data needs to be reduced during the construction of the driving style classifier. Therefore, parameters with high correlation in historical driving data are first removed to obtain driving style feature data. In this embodiment, the selected driving style feature data includes [v e a e [S, Δv, ξ], where v e Indicates the longitudinal speed of the main vehicle, a e ξ represents the longitudinal acceleration of the main vehicle, S represents the relative distance, Δv represents the relative velocity, and ξ represents the actual headway.

[0073] After obtaining the driving style feature data, a driving style classifier can be constructed using this data. To automatically adjust driving modes, this embodiment employs a bidirectional long short-term memory (LSTM) network to build the classifier. The LSM network considers information before and after time t in the input sequence, including an input layer, forward layer, backward layer, fully connected layer, softmax layer, and output layer. The input sequence is typically a 5-10 second time series, and the output is the driver's style, with 1, 2, and 3 representing cautious, normal, and aggressive driving modes, respectively. By selecting specific feature data, the amount of data can be reduced, thereby improving computational efficiency and lowering the cost of training the model.

[0074] In step 5, the driving style classifier is trained and tested multiple times using the training and test sets to obtain the driving style recognition model. In step 6, real-time driving data under following conditions is input into the driving style recognition model to obtain the current driver's driving style. In this way, the appropriate driving mode can be adjusted according to the driver's style to match the driving mode with the driver's style.

[0075] Using Gaussian processes, we define the mean function and covariance function. f(x) = GP(x(x), k(x, z)) represents a Gaussian process, where x and z represent the inputs, x(x) represents the mean function, and k(x, z) represents the covariance function. This represents the variance of the signal controlling the output amplitude. This represents the variance of the signal controlling the input amplitude. To simplify the calculation process, it is usually assumed that x(x) is zero.

[0076] Based on Gaussian processes, hyperparameter learning of the prediction model is performed, assuming a range from x0 to x... n At time t, the input is x = {x0, x1, ..., x2} n The vehicle's motion state is y = {y0, y1, ..., y}. n If the predicted input is x′ and the predicted motion state is y′, then y follows a multivariate Gaussian distribution, described as follows: Where k(x, x) is the covariance function with respect to the training set, k(x, x′) is the covariance function between the training and test sets, and k(x′, x′) is the covariance function with respect to the test set; N() represents a Gaussian distribution with a mean of 0. I represents the environmental noise parameter, and I represents the identity matrix.

[0077] In step 9, according to the formula P(y′|x,x′,y)=N(y′) mean y′ cov ) Calculate the marginal distribution of y′, where, y′mean This represents the prediction result, y′ cov Let represent the covariance of the predicted values, and represent the confidence interval used to reflect the uncertainty of the prediction results. Then, the expression for the uncertainty of the prediction results is: The above derivation yields three hyperparameters. It needs to be optimized.

[0078] In step 10, the hyperparameters of the Gaussian process are learned and optimized based on the training and test sets. The hyperparameter learning results are obtained using a gradient-based optimization algorithm based on the log-likelihood function and its partial derivatives with respect to each hyperparameter. In step 11, based on the hyperparameter learning results and real-time driving data of the vehicle and the preceding vehicle, a vehicle motion state predictor is established. This predicts the motion state of the preceding vehicle, including at least predictions of velocity and acceleration.

[0079] It should be noted that using gradient optimization algorithms to learn hyperparameters is only one implementation method in this embodiment and is not intended to limit the invention. Other algorithms may also be used in other embodiments.

[0080] After obtaining the predicted motion state of the vehicle in front and the driving mode of the vehicle itself, a stochastic model predictive controller needs to be constructed. First, the state equations are established in step 12. These state equations can be established based on the kinematic relationship between the two vehicles, and their expressions are m(t+1)=A m(t)+B u(t)+Gω(t), n(t)=C m(t), where m represents the state variables, m=[Δd, Δv, a e ,j]′,Δd=Sd des Δd represents the distance error, S represents the relative distance, Δv represents the relative speed, and a e Let j represent the longitudinal acceleration of the main vehicle, j represent the impact of the main vehicle, n represent the output variable, u represent the control variable, ω represent the disturbance caused by the motion state of the preceding vehicle, A is the system matrix, B is the control matrix, C is the output matrix, and G is the noise figure matrix.

[0081] Following another vehicle should provide good safety and comfort, while also meeting the driver's needs, such as driver dynamic characteristics. To achieve these requirements, this embodiment designs control objectives based on safety, comfort, and driver dynamic characteristics, such as comfort. in, The weighting coefficient representing acceleration, γ j Weighting coefficients representing impact intensity; driver dynamic characteristics J d =γ Δd Δd 2 +γ Δv Δv 2 , where γΔd γ represents the weighting coefficient of the distance error. Δv Weighting coefficients representing relative velocity; safety Among them, t TTC Indicates the collision time, d safe Indicates the safe following distance, d s Indicates the maximum safe distance.

[0082] Based on the above requirements for following other vehicles, an optimization objective function is constructed, and its expression is as follows: Where st represents a constraint, J total This represents the total cost function.

[0083] In actual driving, the driving state of the vehicle in front is random and the disturbance it causes to the main vehicle is unbounded. Traditional Model Predictive Control (MPC) has certain limitations in dealing with external disturbances. Therefore, this embodiment introduces the probability distribution of the motion state of the vehicle in front under the MPC framework. In step 13, the prediction results are constrained to the state variables in the form of probability constraints. In this way, even if there are external disturbances, the optimal control variables can still be obtained.

[0084] In this embodiment, the probability constraint expression is P(A s m≤b s )≥β, where A s Denotes the state constraint matrix, b s β represents the range of constraints, and β represents the risk probability factor, which represents the minimum probability that the system will satisfy the constraints under random disturbances. The conservatism of the system can be adjusted by adjusting the value of β. The larger the value of β, the more conservative the system.

[0085] By employing a stochastic model predictive control method based on deterministic equivalence to tighten the state boundary, the state variables of the control system are divided into a deterministic part and a disturbance part, expressed as follows: Let represent the deterministic part of the system, and e(t) represent the perturbation part, which includes disturbances caused by the motion of the preceding vehicle and other unknown disturbances. Considering the influence of the perturbation part e(t), the probabilistic constraint is transformed into a deterministic constraint. ε(t) is the equivalent binding force of the probability constraint, erf -1 () is the inverse error function.

[0086] By employing feedback control gain to control the deterministic portion and using model predictive control algorithms to compensate for the disturbance portion, the control variables are obtained. Where g represents the control input provided by the model predictive control algorithm, and Klqr This represents the feedback control gain that enables the system (A, B) to stabilize.

[0087] By constraining the state variables through probabilistic constraints on disturbances generated by the preceding vehicle, and transforming these probabilistic constraints into deterministic constraints, the optimal control variables are obtained using the MPC framework. This ensures that the controller's feasible region remains within a reasonable range, thereby improving the controller's tracking accuracy and comfort. After obtaining the control variables, in step 14, the control variables are input to the main vehicle's underlying controller. The underlying controller converts the control variables into brake / throttle opening positions to achieve vehicle control.

[0088] The anthropomorphic following driving method based on stochastic model predictive control provided in this invention considers the driver's following driving characteristics and the uncertainty of the preceding vehicle's motion. Through an online driving style recognition model, the following process can automatically match the driving mode according to the current driver's style. To address the impact of unbounded disturbances caused by the random motion of the preceding vehicle on the performance of the main vehicle, a Gaussian process is used to predict the uncertainty of the preceding vehicle's motion in real time. The prediction results are then used to constrain the state variables in real time through probabilistic constraints, ensuring that the feasible region of the controller is always kept within a reasonable range. This improves the controller's tracking accuracy and comfort, achieving safe, stable, and comfortable following driving, thereby enhancing the anthropomorphism and acceptability of intelligent vehicles.

[0089] Alternatively, please refer to Figure 2 The following condition includes at least lane restrictions, distance restrictions, speed restrictions, and time restrictions. Lane restrictions include: the lead vehicle and the vehicle in front are traveling in the same lane, and the following condition ends when another vehicle cuts into or out of the space between the two vehicles. Distance restrictions include: the relative distance between the two vehicles is S, where 5m ≤ S ≤ 120m. Speed ​​restrictions include: the speed of the two vehicles is v, where 20km / h ≤ v ≤ 120km / h. Time restrictions include: the following duration is greater than 30s.

[0090] For details, please refer to Figure 2 In the following scenario, both the driving environment and natural driving information need to meet certain requirements, such as lane restrictions, distance restrictions, speed restrictions, and time restrictions. In this embodiment, the lane restriction requires the lead vehicle and the preceding vehicle to travel in the same lane. The following scenario ends when another vehicle cuts into or out of the lane between the two vehicles. The distance restriction is that the relative distance between the two vehicles is S, where 5m ≤ S ≤ 120m. Assuming the speed of both vehicles is v, the speed restriction is 20km / h ≤ v ≤ 120km / h. The time restriction is that the following duration must be greater than 30 seconds.

[0091] Optionally, Figure 3 A flowchart illustrating the selection of driving style feature data provided in this embodiment of the invention is shown below. Figure 1 and Figure 3 In step 4, driving style feature data is selected from historical driving data. Specifically: Step 41, correlation analysis is performed on the historical driving data to obtain correlation values; Step 42, data with correlation values ​​greater than a threshold are removed to obtain driving style feature data as [v e a e [S, Δv, ξ], where v e Indicates the longitudinal speed of the main vehicle, a e ξ represents the longitudinal acceleration of the main vehicle, S represents the relative distance, Δv represents the relative velocity, and ξ represents the actual headway.

[0092] Specifically, commonly used data features representing vehicle-following conditions include the longitudinal speed v of the main vehicle. e longitudinal acceleration a of the main vehicle e Longitudinal displacement of the main vehicle x e Lateral displacement y of the main vehicle e Actual headway ξ and longitudinal speed v of the vehicle in front p Longitudinal displacement of the front vehicle x p Lateral displacement y of the front vehicle p The relative distance S between the main vehicle and the vehicle in front, and the relative speed Δv between the main vehicle and the vehicle in front.

[0093] Please refer to Figure 1 and Figure 3 To improve computational efficiency and reduce the cost of later model training, correlation analysis is performed on the above feature parameters, and feature parameters with high correlation are removed. For example, assuming a correlation threshold of 'a', parameters with a correlation exceeding 'a' are removed, and the remaining parameters can be used as driving style feature parameters. In this embodiment, the final selected driving style feature parameters include the longitudinal velocity of the main vehicle, the longitudinal acceleration of the main vehicle, the relative distance S, and the relative velocity Δv, expressed as [v...]. e a e [S, Δv, ξ]. Where, the relative distance S = x p -x e The actual headway ξ = S / v e Relative velocity Δv = v e -v p .

[0094] Optionally, the hyperparameter learning results can be obtained based on the log-likelihood function and the partial derivatives of each hyperparameter. Specifically, the maximum likelihood method can be used to obtain the hyperparameter learning results through the formula... Calculate the hyperparameters of the Gaussian process, where L(x, ψ) represents the log-likelihood function.

[0095] Specifically, the hyperparameters of the Gaussian process are learned based on the training set. The parameter learning results can be obtained using a gradient-based optimization algorithm based on the log-likelihood function and its partial derivatives with respect to each unknown parameter. The expression for the log-likelihood function is as follows: Using the maximum likelihood method, substitute the log-likelihood function into the formula. The hyperparameters of the Gaussian process can then be obtained.

[0096] Optionally, the disturbance part in, This indicates the determined portion of the disturbance. This represents the probability of the disturbance.

[0097] Specifically, the disturbance generated by the vehicle in front can be decomposed into a disturbance determination part. and perturbation probability part That is to say Assuming the perturbation probability part Follows a Gaussian distribution. Let represent the covariance of the Gaussian distribution. Substituting the decomposition of the perturbation probability part into the state equation, we obtain a new state equation. in, initial state e(0) = 0. Because Then e(t) = N(0, ∑ e，t ), e(t+1)=N(0,∑ e，t+1 And the covariance satisfies ∑ e，t =0, This can be obtained through a vehicle motion state predictor.

[0098] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0099] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A human-like following driving method based on stochastic model predictive control, characterized by, include: Collect historical driving data under following conditions, calibrate the historical driving data, and obtain various calibration data and driver styles; The driving styles include at least the cautious driver, the average driver, and the aggressive driver; The calibration data was used to create training and testing sets; The following model is fitted using the training set and the test set to obtain a driving mode corresponding to the driver's style; the following model is d. des =τv e +d0, where d des v represents the driver's desired distance from other vehicles. e The driving modes include at least a cautious driving mode, a normal driving mode, and an aggressive driving mode. The driving modes represent the longitudinal speed of the main vehicle, d0 represents the minimum safe distance, and τ represents the driver's desired headway. Select driving style feature data from the historical driving data, and use the driving style feature data to construct a driving style classifier; The driving style classifier is trained and tested using the training set and the test set to obtain a driving style recognition model; Real-time driving data under following conditions is collected, and the real-time driving data is input into the driving style recognition model to obtain the driver style. The driving mode is then adjusted according to the driver style to obtain the final driving mode. The mean function and covariance function are defined using a Gaussian process f(x) = GP(x(x), k(x, z)); where f(x) represents a Gaussian process, x and z represent the inputs, x(x) represents the mean function, and k(x, z) represents the covariance function. Assume from x0 to x n At time t, the input is x = {x0, x1, ..., x...} n The vehicle's motion state is y = {y0, y1, ..., y2}. n If the predicted input is x′ and the predicted motion state is y′, then y follows a multivariate Gaussian distribution, described as follows: Where k(x, x) is the covariance function with respect to the training set, k(x, x′) is the covariance function between the training and test sets, and k(x′, x′) is the covariance function with respect to the test set; N() represents a Gaussian distribution with a mean of 0. This represents environmental noise parameters, and I represents the identity matrix. Calculate the marginal distribution of y′, and obtain the expression for the uncertainty of the prediction result as follows: Where, y′ mean This represents the prediction result, y′ cov This represents the covariance of the predicted values; The hyperparameters of the Gaussian process are learned and optimized based on the training set and the test set. The hyperparameter learning results are obtained according to the log-likelihood function and the partial derivatives of each hyperparameter. Based on the learning results, a vehicle motion state predictor is established to obtain a prediction result of the motion state of the vehicle in front. The prediction result of the motion state of the vehicle in front includes at least a prediction result of speed and a prediction result of acceleration. Establish state equations and optimize objective functions to ensure that the following vehicle driving process meets the needs of comfort, safety, and driver requirements; The prediction results are then constrained using probabilistic constraints to obtain control variables; The vehicle is controlled based on the control variables.

2. The humanized car-following driving method based on stochastic model predictive control according to claim 1, characterized in that, The training and test sets are created using the calibration data, specifically as follows: The calibration data is divided into a training set and a test set according to a predetermined ratio.

3. The humanized car following driving method based on stochastic model predictive control according to claim 1, characterized in that, The following conditions include at least lane restrictions, distance restrictions, speed restrictions, and time restrictions; Among them, lane restrictions include: when the main vehicle and the vehicle in front are traveling in the same lane, the following condition ends when another vehicle cuts into or out of the space between the two vehicles; Distance restrictions include: the relative distance between the two vehicles is S, and 5m ≤ S ≤ 120m; The speed limits include: for both vehicles, the speed is v, and 20km / h ≤ v ≤ 120km / h; Time restrictions include: following the vehicle for more than 30 seconds.

4. The humanized car following driving method based on stochastic model predictive control according to claim 1, characterized in that, Driving style characteristic data is selected from the historical driving data, specifically as follows: Correlation analysis was performed on the historical driving data to obtain correlation values; The data with the correlation value greater than the threshold value is removed to obtain driving style feature data as [v e , a e , S, Δv, ξ], wherein v e represents the longitudinal speed of the host vehicle, a e represents the longitudinal acceleration of the host vehicle, S represents the relative distance, Δv represents the relative speed, and ξ represents the actual time headway.

5. The humanized car following driving method based on stochastic model predictive control according to claim 1, characterized in that, the signal variance representing the control output amplitude, the signal variance representing the control input amplitude; the hyperparameters are 6. The humanized car following driving method based on stochastic model predictive control according to claim 5, characterized in that, The marginal distribution of y′ is: P(y′|x,x′,y)=N(y′) mean y′ cov ),in, 7. The anthropomorphic following driving method based on stochastic model predictive control according to claim 6, characterized in that, Based on the log-likelihood function and the partial derivatives of each hyperparameter, the hyperparameter learning results are obtained, specifically: Using the maximum likelihood method, through the formula Calculate the hyperparameters of the Gaussian process, where L(x, ψ) represents the log-likelihood function.

8. The human-like car-following driving method based on stochastic model predictive control according to claim 1, wherein, The state equations are m(t+1)=A m(t)+B u(t)+Gω(t), n(t)=C m(t), where m represents the state variables, m=[Δd, Δv, a e ,j]′,Δd=Sd des Δd represents the distance error, Δv represents the relative speed, and a e Let j represent the longitudinal acceleration of the main vehicle, j represent the impact of the main vehicle, n represent the output variable, u represent the control variable, ω represent the disturbance caused by the motion state of the preceding vehicle, A is the system matrix, B is the control matrix, C is the output matrix, and G is the noise figure matrix. The comfort expression is wherein, a weight coefficient representing acceleration, γ j a weight coefficient representing impact degree; Driver requirements include driver dynamic characteristics, which are expressed as J. d =γ Δd Δd 2 +γ Δv Δv 2 , where γ Δd γ represents the weighting coefficient of the distance error. Δv Weighting coefficients representing relative velocity; The expression of safety is where t TTC represents the collision time, d safe represents the safe following distance, d s represents the limit safety distance; The optimization objective function expression is where s·t· represents a constraint, J total represents the total cost cost function.

9. The humanized car following driving method based on stochastic model predictive control according to claim 8, characterized in that, The prediction results are then constrained using probabilistic constraints to obtain control variables, specifically: By using probability constraints P(A) s m≤b s The state variables are bound together with ≥β, where A s Denotes the state constraint matrix, b s The constraint range is represented by β, and the risk probability factor is represented by β; the state variable expression is: Let represent the deterministic part of the system, and e(t) represent the perturbation part of the system; The determined part is controlled by a feedback control gain, a model predictive control algorithm is used to compensate the disturbance part, and a control variable is obtained wherein g represents a control input provided by the model predictive control algorithm, K lqr represents a feedback control gain that can stabilize the system (A, B).

10. The humanized car following driving method based on stochastic model predictive control according to claim 8, characterized in that, Disturbances caused by the motion state of the preceding vehicle wherein, denotes a disturbance determination part, denotes a disturbance probability part.

Images