Weight self-evolution intelligent automobile personalized lane changing decision-making system

Abstract

The invention relates to a weight self-evolution intelligent automobile personalized lane change decision-making system. The system comprises a driving style quantification module, a lane change decision-making module and a lane change performance baseline evaluation module. The driving style quantification module gives a driving style quantification result and outputs the result to the lane change decision module, the lane change decision module calculates revenue functions of SV and PV and weights of the revenue function of PV and makes lane change decision judgment, and the lane change performance baseline evaluation module combines a decision result output by the data-driven lane change decision sub-module to evaluate the lane change performance baseline. And evaluating a current better decision result, outputting an evaluation result to a weight self-learning sub-module and a baseline self-evaluator, dynamically adjusting each weight of a revenue function of the PV by the weight self-learning sub-module, and determining whether the current weight self-evolution process is ended or not by the baseline self-evaluator according to the evaluation result. The method has an autonomous learning capability, greatly improves the working condition adaptability and flexibility of a lane changing decision system, and can achieve good performance in various complex and changeable driving working conditions.

Description

Technical Field

[0001] The present invention relates to a lane-changing behavior decision system for an autonomous driving vehicle, and in particular to a weight self-evolving intelligent vehicle personalized lane-changing decision system. Background Art

[0002] As the level of automobile intelligence continues to improve, the production and sales of smart cars equipped with high-level autonomous driving systems continue to increase. High-level autonomous driving systems should have the ability to perceive complex traffic situations and make correct behavioral decisions in interactions with surrounding vehicles, thereby reducing the risk of collision and improving traffic efficiency. Lane changing behavior is a core component of microscopic traffic flow and a typical vehicle-to-vehicle interaction behavior. During the lane changing interaction process, whether the vehicle can correctly understand the intentions of surrounding vehicles and make correct decisions is a major technical challenge in the current field of intelligent vehicle technology.

[0003] In order to characterize the personalized differences in driving behaviors of different drivers, a personalized lane-changing decision system for intelligent vehicles combined with a driving style classification algorithm based on driving behavior was established. At present, personalized lane-changing decision systems for intelligent vehicles usually include two categories: data-driven lane-changing decisions and lane-changing decisions based on mechanism rules. Data-driven lane-changing decisions usually use learning methods to take features such as the kinematic parameters of the vehicle and surrounding vehicles as inputs to the learning algorithm, and use neural networks and other methods to fit the intrinsic mapping relationship between the input features and the vehicle's behavioral intentions; lane-changing decisions based on mechanism rules usually use artificially set rules and common sense cognition to establish a mapping relationship between the driver's driving behavior and the driver's intentions. However, data-driven and mechanism-rule-based intelligent vehicle lane-changing behavior decision systems still have the following technical problems:

[0004] 1. The current driving style classification algorithm based on driving behavior can only perform qualitative analysis, but it is difficult to perform quantitative analysis. The currently constructed driving style classification algorithm only divides the driver's driving style into three categories: cautious, general, and aggressive. In fact, there are individual differences in drivers' driving styles, and they are easily affected by the traffic environment. Accurately evaluating driving style is particularly important for correctly judging the driver's driving intention and making correct decisions. However, the existing driving style classification algorithm is difficult to provide quantitative indicators of driving style.

[0005] 2. The data-driven lane-changing decision system establishes a multi-dimensional mapping relationship between the kinematic variables of the surrounding vehicles and the vehicle itself during the driving process and the lane-changing decision of the vehicle itself. Its essence is to establish an endogenous relationship between the input parameters and the output results by learning the implicit interaction logic in massive data. Since the data-driven method belongs to the black-box algorithm, the algorithm has poor interpretability, and it is difficult to make an explicit expression of the interaction between each system parameter and the traffic vehicle during the lane-changing decision. This leads to practical difficulties in adjusting the parameters of the system, which further leads to unstable system output and even safety hazards.

[0006] 3. The lane-changing decision system based on mechanism rules uses the observation of the driver's driving behavior to establish a functional mapping relationship with practical meaning between the vehicle's kinematic information and the vehicle's driving intention, comprehensively considers the vehicle's kinematic constraints and driving safety, and makes lane-changing decisions based on manually set decision rules. However, the actual traffic environment is complex and changeable, and it is difficult for artificial rules to cover all kinds of driving conditions, which causes the problem of poor versatility of the lane-changing decision system based on mechanism rules. At the same time, the parameters in the lane-changing decision system based on mechanism rules are usually fixed, which makes it difficult for the system to have the ability to learn autonomously, further limiting the performance of the lane-changing decision system based on mechanism rules. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a weight self-evolving intelligent automobile personalized lane changing decision system, comprising: a driving style quantification module, a lane changing decision module, and a lane changing performance baseline evaluation module; the driving style quantification module comprises a kinematic feature extraction submodule, a driving style evaluation submodule and a driving style quantification data temporary storage submodule; the lane changing decision module comprises: a lane changing benefit function submodule, a decision judgment submodule, a weight self-learning submodule and a driving style and dynamic weight storage submodule; the lane changing performance baseline evaluation module comprises a data-driven lane changing decision submodule, a decision evaluation submodule and a baseline self-evaluator.

[0008] Furthermore, the working process of the driving style quantification module is as follows: first, the kinematic information of the vehicle behind the target lane is transmitted to the kinematic feature extraction submodule, the kinematic feature extraction submodule performs data preprocessing and data feature extraction on the transmitted kinematic information, and then transmits the processed information to the driving style evaluation submodule, the trained Gaussian mixture model (abbreviated as GMM) in the driving style evaluation submodule receives the above information, and gives the quantified probability of each driving style of the vehicle behind the target lane according to the obtained information, including: the probability of cautious type x d1 , the general probability x d2 , the probability of radical type x d3 , and the probability x d1 、x d2 、x d3Output to the driving style quantization data temporary storage submodule, which organizes the data into a vector X with one row and three columns d =[x d1 ,x d2 ,x d3 ], i.e., driving style quantization data, and is stored in a data register; when new driving style quantization data is transmitted to the driving style quantization data temporary storage submodule, the previous data stored in the data register will be overwritten by the new quantization data.

[0009] Specifically, the data feature extraction function of the kinematic feature extraction submodule is to obtain the relative speed, relative acceleration and jerk information of the vehicle behind the target lane relative to the vehicle within the time T, calculate the average relative speed, average relative acceleration and average jerk, and pass the calculated average relative speed, average relative acceleration and average jerk information to the driving style evaluation submodule.

[0010] Specifically, the training process of the Gaussian mixture model is as follows:

[0011] Step 1: Preprocessing and feature extraction: Obtain the relative speed, relative acceleration and jerk information of the target lane rear vehicle with N lane-changing behaviors in the data set within T time before the ego vehicle changes lanes, and calculate the average relative speed of the target lane rear vehicle in the i-th training data. Average relative acceleration and average jerk Recorded as All N qualified training data are expressed as: P = [P1, P2, ..., P n ];

[0012] Step 2: Parameter training of Gaussian mixture model: All N qualified training data are passed into the Gaussian mixture model. The number of sub-Gaussian models K in the Gaussian mixture model is set to 3. The maximum expectation algorithm is used to solve the parameters of the Gaussian mixture model. It is mainly divided into two steps:

[0013] The first step, E-step, calculates the Q function according to the following formula, which represents the probability that the i-th data belongs to the k-th sub-Gaussian distribution:

[0014]

[0015] Among them, φ k represents the probability of the kth sub-Gaussian distribution, x (i) represents the i-th data, μ k represents the mean of the kth sub-Gaussian distribution, σ k Represents the standard deviation of the k-th sub-Gaussian distribution, and φ is randomly given in the first iteration k , μk , σ k The value of

[0016] The second step, M-step, updates the model parameters according to the following formula:

[0017]

[0018] Repeat the E-step and M-step iterations several times until the model converges.

[0019] After the training is completed, the Gaussian mixture model evaluates the probability x that the driving style of the following vehicle is cautious by inputting the average relative speed, average relative acceleration, and average jerk information of the following vehicle within the time T before the self-vehicle changes lanes. d1 , the general probability x d2 , the probability of radical type x d3 , and output the above probability as the output of the driving style evaluation submodule.

[0020] Furthermore, the working process of the lane change decision module is as follows: kinematic information of the vehicle behind the target lane, kinematic information of the vehicle itself, kinematic information of the surrounding vehicles, and driving style quantification data X of the driving style quantification module are combined. d =[x d1 ,x d2 ,x d3 ] is passed to the lane-changing benefit function submodule, which is based on the established benefit function of the ego vehicle SV and the benefit function of the vehicle behind the target lane PV; first query whether there is a dynamic weight record of the same driving style quantified data in the driving style and dynamic weight storage submodule, if there is a record, then call the dynamic weight record as the weights ω1 and ω2 of the PV benefit function; if there is no record, then according to the driving style quantified data X d =[x d1 ,x d2 ,x d3 ]Preliminary adjustment of the weights of the PV return function: Then, the SV benefit function result is calculated based on the established benefit function of the ego vehicle SV, and the PV benefit function result is calculated based on the established benefit function of the vehicle behind the target lane PV; the calculated SV benefit function result, PV benefit function result and each weight of the PV benefit function are transmitted to the decision judgment submodule, and the decision judgment submodule first determines the lane change benefit R of the SV at this time lane Is it greater than the lane keeping benefit R of staying in the original lane? stay , if the lane-changing benefit of SV is R lane No greater than the lane keeping benefit R of staying in the original lane stay , then the decision of the vehicle SV to keep the lane is given; if the lane change benefit Rlane Greater than the lane keeping benefit R of staying in the original lane stay , the decision-making submodule is based on the game idea that PV expects to obtain the maximum benefit. It defaults that PV will take the action of evading and slowing down or continuing to drive without evading according to the principle of maximizing its own benefits. Based on this, the decision of whether the self-vehicle SV should change lanes or keep lanes is given. Then the decision is passed to the decision-making evaluation submodule of the lane-changing performance baseline evaluation module, and the weights of the PV benefit function are passed to the weight self-learning submodule. The decision-making evaluation submodule determines whether the weights of the current PV benefit function have optimization space, and outputs the evaluation results to the weight self-learning submodule. The weight self-learning submodule receives the PV benefits from the decision-making submodule. The weights of the function, the evaluation results from the decision evaluation submodule and the driving style quantification data from the driving style quantification data temporary storage submodule. If the evaluation results from the decision evaluation submodule show that there is room for optimization in the weights of the PV benefit function, the weight self-learning submodule then guides the dynamic adjustment of the weights of the PV benefit function, and stores the new weights and the corresponding driving style quantification data into the driving style and dynamic weight storage submodule. Otherwise, the weight self-learning submodule does not adjust the weights of the PV benefit function, and directly passes the corresponding weights and the corresponding driving style quantification data to the driving style and dynamic weight storage submodule.

[0021] Specifically, at the current time step t, the lane change benefit function submodule calculates the SV benefit function result and PV benefit function result of the current time step t based on the kinematic information of the target lane rear vehicle at the current time step t obtained from the data set, the kinematic information of the ego vehicle, the kinematic information of the surrounding vehicles, and the current driving style quantification data obtained from the driving style quantification module, using the established ego vehicle SV benefit function and PV benefit function. The established ego vehicle SV benefit function and PV benefit function are as follows:

[0022] SV return function S priceSV for:

[0023]

[0024] In the above formula, K SV1 , K SV2 , K SV3 , Both are positive proportional coefficients, R lane is the lane-changing benefit of the SV when making a lane-changing decision;

[0025] V S (v SV ,v FV1 ) function represents the desired benefit of the ego vehicle SV to ensure its own driving speed by changing lanes, which is expressed as: Among them, v SVis the velocity of the ego vehicle SV at the current time step t, v FV1 is the velocity of FV1 at the current time step t;

[0026] V FV (v FV1 ,v FV2 ) function represents the benefit of the vehicle’s lane-changing action in improving its subsequent following speed, which can be expressed as: Among them, v FV1 is the velocity of FV1 at the current time step t, v FV2 is the velocity of FV2 at the current time step t;

[0027] S PV (S SV ,S PV ) function represents the safety benefit based on the longitudinal distance between the self-vehicle SV and the rear vehicle PV, which is expressed as:

[0028]

[0029] Among them, S SV is the longitudinal position of SV at the current time step t (with the vehicle's travel direction as the positive direction of the coordinate system), S PV is the longitudinal position of PV at the current time step t; S PV (S SV ,S PV ) is a piecewise function. When the longitudinal distance S between SV and PV is SV -S PV Exceeding the safety threshold S s , providing a constant positive return R sp ; When the longitudinal distance S between SV and PV SV -S PV At the safety threshold S s With the danger threshold S d When the spacing is between, it provides a positive return that gradually decreases as the spacing decreases. Until the risk threshold is reached, the return drops to 0; when the vertical distance S between SV and PV SV -S PV Below the danger threshold S d When the spd ;

[0030] Function representation The SV behavior decision constraint function based on the longitudinal distance between SV and FV1 and the longitudinal speed is expressed as: Among them, S FV1 is the longitudinal position of FV1 at the current time step t. When the longitudinal arrival time of SV and FV1 is When it is less than 3.5 seconds,

[0031] Provides a large negative return R spn1 ; When the longitudinal arrival time of SV and FV1 When it is greater than or equal to 3.5 seconds, Provides a small positive return R sp1 ;

[0032] Function representation The SV behavior decision constraint function based on the longitudinal distance between SV and FV2 and the longitudinal speed is expressed as: Among them, S FV2 is the longitudinal position of FV2 at the current time step t. When the longitudinal arrival time of SV and FV2 is Less than the minimum lane-changing safety following time difference t sp2 =3.5+t lanec hour, Provides a large negative return R spn2 ; When the longitudinal arrival time of SV and FV2 Greater than or equal to the minimum lane-changing safety following time difference t sp2 hour,

[0033] Provides a small positive return R sp2 , where t lanec is the average lane changing time of the driver, t lanec The average lane changing time of drivers is obtained by counting the public data set;

[0034] When the SV does not change lanes and stays in the original lane, it obtains a constant small negative benefit R stay .

[0035] PV return function S pricePV for:

[0036] S pricePV =ω1·V PV (a PV ,v PV )+ω2·S Space (S PV ,S SV )

[0037] In the above formula, ω1 and ω2 represent the weights of the PV return function and are dynamically variable positive weight coefficients. PV (a PV ,v PV ) function represents the speed requirement benefit of PV, expressed as: Among them, v PV is the velocity of PV at the current time step t, a PVis the acceleration value of PV at the current time step t. When the action of PV is to avoid deceleration, a PV is a negative constant. When the PV action is not to avoid, a PV =0, τ0 is a positive constant; S Space (S PV ,S SV ) function represents the required return of PV, expressed as: Among them, h0 is a positive constant, S SV is the longitudinal position of SV at the current time step t, S PV is the longitudinal position of PV at the current time step t.

[0038] Specifically, the decision-making process of the decision judgment submodule is as follows: the decision judgment submodule obtains the SV benefit function result and the PV benefit function result of the current time step t from the lane change benefit function submodule, and first determines the lane change benefit R of the SV at this time. lane Is it greater than the lane keeping benefit R of staying in the original lane? stay , if the lane-changing benefit of SV is R lane No greater than the lane keeping benefit R of staying in the original lane stay , the decision-making submodule determines that the lane-changing prerequisite is not met at this time, and outputs the decision that the SV should keep the lane; if the lane-changing benefit R of SV at this time lane Greater than the lane keeping benefit R of staying in the original lane stay , the decision judgment submodule determines that the prerequisite for lane changing is met at this time; then the PV avoidance deceleration benefit and the PV non-avoidance benefit are calculated, and the benefit values ​​of the two are compared, and it is considered that PV will take an action with a larger benefit value: if the PV avoidance deceleration benefit is greater than or equal to the non-avoidance benefit, then it is judged that PV takes an avoidance deceleration action; if the PV avoidance deceleration benefit is less than the non-avoidance benefit, then it is judged that PV takes a non-avoidance action, and based on this, a decision is given as to whether the ego vehicle SV should change lanes or continue lane keeping at this time: if the PV action is judged to be avoidance deceleration, then the output decision of the decision judgment submodule at this time is lane changing; if the PV action is judged to be non-avoidance, then the output decision of the decision judgment submodule at this time is lane keeping, and then the decision judgment submodule passes the decision to the decision evaluation submodule of the lane changing performance baseline evaluation module, and passes the weights of the PV benefit function to the weight self-learning submodule.

[0039] Furthermore, the working process of the lane changing performance baseline evaluation module is as follows: first, the kinematic information of the target lane rear vehicle, the kinematic information of the vehicle itself, the kinematic information of the surrounding vehicles, and the driving style quantification data of the driving style quantification module extracted from the data set are used as the current state, and after data preprocessing, they are transmitted to the trained data-driven lane changing decision submodule. The trained data-driven lane changing decision submodule outputs the action of the current SV, i.e., lane changing or lane keeping, based on the current state transmitted, and then transmits the action to the decision evaluation submodule. The decision evaluation submodule determines the lane changing decision submodule based on the decision judgment submodule of the lane changing decision module. The decision of whether the SV should change lanes or keep lane, which is passed into the module, is compared with the output of the data-driven lane changing decision submodule, and the output decision result of the current data-driven lane changing decision submodule and the output decision result of the decision judgment submodule of the lane changing decision module are evaluated. Then the evaluation result is output to the baseline self-evaluator and the weight self-learning submodule of the lane changing decision module. The weight self-learning submodule dynamically adjusts the weights of the PV benefit function according to the evaluation result of the decision evaluation submodule. The baseline self-evaluator decides whether the current weight self-evolution process can be terminated according to the evaluation result of the decision evaluation submodule.

[0040] Specifically, the data preprocessing process before being passed into the trained data-driven lane-changing decision submodule is as follows:

[0041] The kinematic information of the target lane vehicle behind, the vehicle kinematic information, the surrounding vehicle kinematic information, and the driving style quantification data X obtained from the driving style quantification module are d Composition state dataset S D :

[0042]

[0043] Among them, i represents that the state is the i-th time series extracted from the data set, 0≤i≤N E ; For the i-th time sequence state sequence S D(i) , expressed as:

[0044]

[0045] Where t represents the current state data as the state data of the tth sampling point, 0≤t≤T E , the i-th time series S D(i,t) The state data at the tth sampling point is expressed as:

[0046] S D(i,t) =[E PV(i,t) ,E SV(i,t) ,E FV1(i,t) ,E FV2(i,t) ,X d(i) ]

[0047] Among them, E PV(i,t) represents the kinematic information of the vehicle PV behind the target lane at the tth sampling point, E SV(i,t) represents the kinematic information of the ego vehicle SV at the tth sampling point, E FV1(i,t) represents the kinematic information of the front vehicle FV1 in the lane before the ego vehicle changes lanes at the tth sampling point, E FV2(i,t) represents the kinematic information of the front vehicle FV2 in the target lane at the tth sampling point, X d(i) Represents the driving style quantification data of the PV of the i-th time series.

[0048] Specifically, the training process of the data-driven lane-changing decision submodule is as follows:

[0049] First, from the state data set S D Take any time sequence S from D(i) , and then the state data of each sampling point of the time series state sequence is sequentially transmitted to the data-driven lane-changing decision submodule in time sequence; at the tth time step, S D(i,t) The deep Q network in the data-driven lane-changing decision submodule calculates the value of the current action state. And make an action t , where action a t is one of two discrete actions: lane keeping or lane changing. is the model parameter of the deep Q network, and then comes the t+1th time step, S D(i,t+1) The deep Q network in the data-driven lane-changing decision submodule obtains the reward function R S (t) Define the true reward r t , based on maximizing the action state value at this time The principle of estimating action a t+1 = argmaxQ(s t+1 ,a,ω), calculate the timing target Where 0<γ<1, which is the minimum loss Loss: Perform gradient descent to update model parameters: Where η is a positive constant;

[0050] Repeat the above steps until the model parameters converge, and the data-driven lane-changing decision submodule is trained. D(i,t) The data-driven lane-changing decision submodule is input, and the data-driven lane-changing decision submodule outputs the decision at time step t.

[0051] Furthermore, the reward function R of the deep Q network of the data-driven lane-changing decision submodule is S(t) is defined as follows:

[0052]

[0053] Among them, μ1, μ2, μ3 are constant coefficients, v SV (t) indicates the SV(i,t) The obtained SV longitudinal velocity information, v FV1 (t) indicates the FV1(i,t) The longitudinal velocity information of FV1 is obtained, v FV2 (t) indicates the FV2(i,t) The longitudinal velocity information of FV2 is obtained, S SV (t) indicates the SV(i,t) The longitudinal position information of the SV obtained, S PV (t) indicates the PV(i,t) The longitudinal position information of the PV obtained, S d is the danger threshold.

[0054] Specifically, the evaluation process of the decision evaluation submodule is as follows:

[0055] The decision evaluation submodule receives the decision action A1 from the decision judgment submodule of the lane change decision module and the decision action A2 from the data-driven lane change decision submodule, and determines the quality of the action and takes action according to the output values ​​of the two actions according to the following judgment rules:

[0056] (1) If A1 = A2 = lane keeping, then at this moment, since lane changing behavior has not occurred, the decision evaluation submodule only outputs decision-consistent information to the baseline self-evaluator.

[0057] (2) If A1 = lane keeping and A2 = lane changing, then the decision evaluation submodule uses the cubic polynomial method to keep the speed at the start and end of the lane change consistent, and the lane change duration is t lanec Plan a lane-changing route and attempt to change lanes; during the lane-changing attempt, the trajectories of the surrounding vehicles and PV are updated based on the kinematic information obtained from the data set, and the SV's trajectory is updated based on the planned lane-changing route: If the longitudinal distance between SV and PV is lower than the safety threshold S during the lane-changing attempt, d , it is considered that lane changing is inferior to lane keeping at this moment, so the decision evaluation submodule only outputs the information that is better than the lane changing decision module to the baseline self-evaluator; if the longitudinal distance between SV and PV is always higher than the safety threshold S during the lane changing attempt, d , it is considered that lane changing is better than lane keeping at this moment, so the decision evaluation submodule outputs the information that lane changing is currently possible to the weight self-learning submodule of the lane changing decision module, and outputs the information of decision inconsistency to the baseline self-evaluator.

[0058] (3) If A1 = lane change, A2 = lane change, then at this moment, since the actions are consistent, the decision evaluation submodule only outputs the information of consistent decisions to the baseline self-evaluator.

[0059] (4) If A1 = lane change, A2 = lane keeping, then the decision evaluation submodule uses a cubic polynomial to plan the lane change route and attempts to change lanes. During the lane change attempt, the trajectories of the surrounding vehicles and PV are updated based on the kinematic information obtained from the data set, and the SV's trajectory is updated based on the planned lane change route. If the longitudinal distance between SV and PV is lower than the safety threshold S during the lane change attempt, d , it is considered that lane changing at this moment is inferior to lane keeping. Therefore, the decision evaluation submodule outputs information about the risk of the current lane change to the weight self-learning submodule of the lane change decision module, and outputs information about inconsistent decisions to the baseline self-evaluator. If the longitudinal distance between SV and PV is always higher than the safety threshold S during the lane change attempt, d , it is considered that lane changing is better than lane keeping at this moment, so the decision evaluation submodule only outputs the better information of the lane changing decision module to the baseline self-evaluator.

[0060] Specifically, the operation process of the baseline self-assessor is as follows:

[0061] The baseline self-evaluator receives the information output by the decision evaluation submodule and counts the number of information h1 that is better than the lane change decision module, the number of information h2 that is inconsistent with the decision, and the number of information h3 that is consistent with the decision every M (M is a positive constant) data. and It is considered that the decision of the decision judgment submodule is significantly better than the decision of the data-driven lane change decision submodule at this time, and the weight self-evolution process is terminated, otherwise the weight self-evolution process continues; β0 and β1 are positive constants greater than zero.

[0062] Specifically, the operation process of the weight self-learning submodule is as follows:

[0063] The weight self-learning submodule receives the weights of the PV benefit function of the decision judgment submodule and the information output by the decision evaluation submodule:

[0064] (1) When receiving information from the decision evaluation submodule that the current lane change is risky, it is considered that the weight of the speed benefit term in the PV benefit function is underestimated at this time. Therefore, the weight self-learning submodule adjusts the PV benefit function S according to the following formula: pricePV The weights ω1 and ω2 are the new weights ω1' and ω'2:

[0065] 1-ω1'→ω'2, where α is a positive constant coefficient.

[0066] (2) When receiving the information that the lane can be changed from the decision evaluation submodule, it is considered that the weight of the distance benefit term in the PV benefit function is underestimated at this time. Therefore, the weight self-learning submodule adjusts the PV benefit function S according to the following formula: pricePV The weights ω1 and ω2 are the new weights ω1' and ω'2:

[0067] 1-ω'2→ω1', where α is a positive constant coefficient.

[0068] (3) When no information is received from the decision evaluation submodule, the weight self-learning submodule does not adjust the weights of the PV benefit function.

[0069] Subsequently, the weight self-learning submodule transfers the latest weights of the PV benefit function and its corresponding driving style quantification data to the driving style and dynamic weight storage submodule for storage.

[0070] Specifically, the operation process of the driving style and dynamic weight storage submodule is as follows:

[0071] First, the driving style and dynamic weight storage submodule receives the weights of the PV benefit function and its corresponding driving style quantitative data newly transmitted from the weight self-learning submodule; then, the driving style and dynamic weight storage submodule queries whether there is a driving style quantitative data record in the current storage that is consistent with the newly transmitted driving style quantitative data, and performs the following operations:

[0072] (1) If there is a driving style quantization data record that is consistent with the newly-input driving style quantization data, the driving style and dynamic weight storage submodule updates the original record, and overwrites the weight records of the original PV benefit function with the weights of the newly-input PV benefit function.

[0073] (2) If there is no driving style quantification data record that is consistent with the newly-input driving style quantification data, the driving style and dynamic weight storage submodule stores the weights of the newly-input PV benefit function and its corresponding driving style quantification data as a record in the storage.

[0074] The overall process of weight self-evolution learning of a weight self-evolution intelligent vehicle personalized lane change decision system provided by the present invention is as follows:

[0075] Step 1: extract the kinematic information of the target lane rear vehicle, the ego vehicle kinematic information and the surrounding vehicle kinematic information from the data set, and then send the above information to the driving style quantification module, the lane change decision module and the lane change performance baseline evaluation module; first, randomly extract a driving information from the data set, and a driving information includes the kinematic information of the target lane rear vehicle, the ego vehicle kinematic information and the surrounding vehicle kinematic information at each time step of a lane change process; at time step t, the information input to the driving style quantification module is the kinematic information of the target lane rear vehicle of the driving information during the period from time step t-T+1 to time step t, a total of T time; at time step t, the information input to the lane change decision module and the lane change performance baseline evaluation module is the kinematic information of the target lane rear vehicle, the ego vehicle kinematic information and the surrounding vehicle kinematic information of the driving information at time step t.

[0076] Step 2: The driving style quantification module first transfers the kinematic information of the vehicle behind the target lane to the kinematic feature extraction submodule. The kinematic feature extraction submodule performs data preprocessing and data feature extraction on the kinematic information, and then transfers the processed information to the driving style evaluation submodule. The trained Gaussian mixture model in the driving style evaluation submodule receives the above information and gives the quantified probability of each driving style of the vehicle behind the target lane based on the information obtained, including: the probability of cautious type x d1 , the general probability x d2 , the probability of radical type x d3 , and the probability x d1 、x d2 、x d3 Output to the driving style quantization data temporary storage submodule, which organizes the data into a vector X with one row and three columns d =[x d1 ,x d2 ,x d3 ], i.e., driving style quantization data, and is stored in a data register; when new driving style quantization data is transmitted to the driving style quantization data temporary storage submodule, the previous data stored in the data register will be overwritten by the new quantization data.

[0077] Step three: the lane changing decision module calculates the benefit functions of SV and PV through the lane changing benefit function submodule based on the kinematic information of the vehicle behind the target lane, the kinematic information of the vehicle itself and the kinematic information of the surrounding vehicles and the driving style quantification data of the driving style quantification module, and outputs the calculation results and the weights of the PV benefit function to the decision judgment submodule; the decision judgment submodule outputs the decision judgment results to the decision evaluation submodule of the lane changing performance baseline evaluation module, and outputs the weights of the PV benefit function to the weight self-learning submodule.

[0078] Step 4: The data-driven lane-changing decision submodule of the lane-changing performance baseline evaluation module outputs the decision result to the decision evaluation submodule based on the kinematic information of the vehicle behind the target lane, the kinematic information of the vehicle itself, the kinematic information of the surrounding vehicles and the driving style quantification data of the driving style quantification module.

[0079] Step 5: The decision evaluation submodule evaluates which one is better, the output decision result of the current data-driven lane changing decision submodule or the output decision result of the decision judgment submodule of the lane changing decision module, and then outputs the evaluation result to the weight self-learning submodule and the baseline self-evaluator of the lane changing decision module.

[0080] Step six: the weight self-learning submodule dynamically adjusts the weights of the PV benefit function according to the evaluation results of the decision evaluation submodule, and transmits the latest weights of the PV benefit function and the corresponding driving style quantification data obtained from the driving style quantification data temporary storage submodule to the driving style and dynamic weight storage submodule of the lane change decision module. The driving style and dynamic weight storage submodule stores the transmitted information.

[0081] In step seven, the baseline self-evaluator decides whether the current weight self-evolution process can be terminated based on the evaluation results of the decision evaluation submodule. If the weight has not been adjusted to the optimal state, steps one to seven are repeated; if the weight has been adjusted to the optimal state, the weight self-evolution process is terminated.

[0082] Beneficial effects of the present invention:

[0083] 1. The weighted self-evolving intelligent vehicle personalized lane-changing decision system of the present invention realizes the abstract representation of the relationship between the kinematic parameters of the rear vehicle and the driving style of the rear vehicle driver by analyzing the kinematic explicit characteristics of the rear vehicle in the adjacent lane when the vehicle has the intention to change lanes, and further establishes a driving style quantitative index, and quantitatively describes the driving style of the rear vehicle driver and the differences between different driving styles through a probability model;

[0084] 2. The weighted self-evolving intelligent car personalized lane-changing decision system described in the present invention deeply considers the behavioral interaction between vehicles during the actual lane-changing process, and establishes a lane-changing decision module based on game theory, so that the system can explicitly express the dynamic and time-varying vehicle-to-vehicle interaction phenomenon during the lane-changing period, thereby making the lane-changing decision logic clear and transparent, and enhancing the interpretability of the algorithm. By comprehensively considering the driving benefits in many aspects, a multi-weighted game benefit function with practical meaning is established, and lane-changing decisions are made based on the benefit function, which improves the stability of the system decision output;

[0085] 3. The weighted self-evolving intelligent automobile personalized lane changing decision system described in the present invention greatly improves the adaptability and flexibility of the lane changing decision system by dynamically adjusting the weight of the game reward function, introduces a lane changing performance baseline evaluation module based on deep reinforcement learning, and forms a coupling connection with the driving style quantification module, thereby establishing a quantitative mapping relationship between driving style and lane changing decision results, and enables the established lane changing decision system to have autonomous learning capabilities, and can achieve good performance in various complex and changeable driving conditions, making it possible to guide the stable improvement of the lane changing decision system through data-driven methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 Schematic diagram of the overall framework of the lane changing decision system of the present invention.

[0087] Figure 2 Schematic diagram of the position relationship between the ego vehicle SV and surrounding traffic vehicles in a typical lane-changing scenario.

[0088] Figure 3 This is a schematic diagram of the driving style quantification module principle of the present invention.

[0089] Figure 4 It is a schematic diagram of the decision-making and judgment submodule flow of the present invention.

[0090] Figure 5 Schematic diagram of the framework of the data-driven lane-changing decision submodule of the present invention.

[0091] Figure 6 It is a schematic diagram of the actual application flow of the decision-making system of the present invention. DETAILED DESCRIPTION

[0092] See also Figure 1 As shown, a weight self-evolving intelligent automobile personalized lane changing decision system provided by the present invention includes three main modules: a driving style quantification module, a lane changing decision module, and a lane changing performance baseline evaluation module; the driving style quantification module includes a kinematic feature extraction submodule, a driving style evaluation submodule and a driving style quantification data temporary storage submodule; the lane changing decision module includes: a lane changing benefit function submodule, a decision judgment submodule, a weight self-learning submodule and a driving style and dynamic weight storage submodule; the lane changing performance baseline evaluation module includes a data-driven lane changing decision submodule, a decision evaluation submodule and a baseline self-evaluator.

[0093] First, a piece of driving information is randomly extracted from the data set. A piece of driving information includes the kinematic information of the target lane rear vehicle, the self-vehicle kinematic information and the surrounding vehicle kinematic information at each time step of a lane changing process; at time step t, the information input to the driving style quantification module is the kinematic information of the target lane rear vehicle during the period T from time step t-T+1 to time step t; at time step t, the information input to the lane changing decision module and the lane changing performance baseline evaluation module is the kinematic information of the target lane rear vehicle, the self-vehicle kinematic information and the surrounding vehicle kinematic information of the driving information at time step t;

[0094] The target lane rear vehicle kinematic information includes the target lane rear vehicle (PV)'s longitudinal and lateral position information, longitudinal and lateral velocity information, longitudinal and lateral acceleration information, jerk information and its relative velocity and relative acceleration information to the ego vehicle (SV); the ego vehicle kinematic information includes the ego vehicle's longitudinal and lateral position information, longitudinal and lateral velocity information and longitudinal and lateral acceleration information; the surrounding vehicle kinematic information includes the longitudinal and lateral position information and longitudinal and lateral velocity information of the front vehicle (FV1) in the lane where the ego vehicle is located before changing lanes and the front vehicle (FV2) in the target lane. The above information is obtained from public data sets.

[0095] Typical lane-changing scenarios include Figure 2 As shown in the figure, SV is the lane-changing vehicle, PV is the vehicle behind in the target lane, SV expects to change lanes from its lane to the lane where PV is located, the front vehicle FV1 in the lane where SV is located before changing lanes and the front vehicle FV2 in the target lane provide kinematic constraints for the lane-changing behavior of SV, and the interactive confrontation process between SV and PV will determine the lane-changing decision of SV.

[0096] like Figure 1 , 3 As shown in FIG. 1 , the working process of the driving style quantification module is as follows: first, the kinematic information of the vehicle behind the target lane is transmitted to the kinematic feature extraction submodule, the kinematic feature extraction submodule performs data preprocessing and data feature extraction on the transmitted kinematic information, and then transmits the processed information to the driving style evaluation submodule, the trained Gaussian mixture model (abbreviated as GMM) in the driving style evaluation submodule receives the above information, and gives the quantified probability of each driving style of the vehicle behind the target lane according to the obtained information, including: the probability of cautious type x d1 , the general probability x d2 , the probability of radical type x d3 , and the probability x d1 、x d2 、x d3 Output to the driving style quantization data temporary storage submodule, which organizes the data into a vector X with one row and three columns d =[x d1 ,x d2,x d3 ], i.e., driving style quantization data, and is stored in a data register; when new driving style quantization data is transmitted to the driving style quantization data temporary storage submodule, the previous data stored in the data register will be overwritten by the new quantization data.

[0097] Specifically, the data feature extraction function of the kinematic feature extraction submodule is to obtain the relative speed, relative acceleration and jerk information of the vehicle behind the target lane relative to the vehicle within the time T, calculate the average relative speed, average relative acceleration and average jerk, and pass the calculated average relative speed, average relative acceleration and average jerk information to the driving style evaluation submodule.

[0098] Specifically, the training process of the Gaussian mixture model is as follows:

[0099] Step 1: Preprocessing and feature extraction: Obtain the relative speed, relative acceleration and jerk information of the target lane rear vehicle with N lane-changing behaviors in the data set within T time before the ego vehicle changes lanes, and calculate the average relative speed of the target lane rear vehicle in the i-th training data. Average relative acceleration and average jerk Recorded as All N qualified training data are expressed as: P = [P1, P2, ..., P n ];

[0100] Step 2: Parameter training of Gaussian mixture model: All N qualified training data are passed into the Gaussian mixture model. The number of sub-Gaussian models K in the Gaussian mixture model is set to 3. The maximum expectation algorithm is used to solve the parameters of the Gaussian mixture model. It is mainly divided into two steps:

[0101] The first step, E-step, calculates the Q function according to the following formula, which represents the probability that the i-th data belongs to the k-th sub-Gaussian distribution:

[0102]

[0103] Among them, φ k represents the probability of the kth sub-Gaussian distribution, x (i) represents the i-th data, μ k represents the mean of the kth sub-Gaussian distribution, σ k Represents the standard deviation of the k-th sub-Gaussian distribution, and φ is randomly given in the first iteration k , μ k , σ k The value of

[0104] The second step, M-step, updates the model parameters according to the following formula:

[0105]

[0106] Repeat the E-step and M-step iterations several times until the model converges.

[0107] After the training is completed, the Gaussian mixture model evaluates the probability x that the driving style of the following vehicle is cautious by inputting the average relative speed, average relative acceleration, and average jerk information of the following vehicle within the time T before the self-vehicle changes lanes. d1 , the general probability x d2 , the probability of radical type x d3 , and output the above probability as the output of the driving style evaluation submodule.

[0108] like Figure 1 , 4 As shown, the working process of the lane change decision module is as follows: the kinematic information of the vehicle behind the target lane, the kinematic information of the vehicle itself, the kinematic information of the surrounding vehicles, and the driving style quantification data X of the driving style quantification module are combined. d =[x d1 ,x d2 ,x d3 ] is passed to the lane-changing benefit function submodule, which is based on the established benefit function of the ego vehicle SV and the benefit function of the vehicle behind the target lane PV; first query whether there is a dynamic weight record of the same driving style quantified data in the driving style and dynamic weight storage submodule, if there is a record, then call the dynamic weight record as the weights ω1 and ω2 of the PV benefit function; if there is no record, then according to the driving style quantified data X d =[x d1 ,x d2 ,x d3 ]Preliminary adjustment of the weights of the PV return function: Then, the SV benefit function result is calculated based on the established benefit function of the ego vehicle SV, and the PV benefit function result is calculated based on the established benefit function of the vehicle behind the target lane PV; the calculated SV benefit function result, PV benefit function result and each weight of the PV benefit function are transmitted to the decision judgment submodule, and the decision judgment submodule first determines the lane change benefit R of the SV at this time lane Is it greater than the lane keeping benefit R of staying in the original lane? stay , if the lane-changing benefit of SV is R lane No greater than the lane keeping benefit R of staying in the original lane stay , then the decision of the vehicle SV to keep the lane is given; if the lane change benefit R lane Greater than the lane keeping benefit R of staying in the original lane stay, the decision-making submodule is based on the game idea that PV expects to obtain the maximum benefit. It defaults that PV will take the action of evading and slowing down or continuing to drive without evading according to the principle of maximizing its own benefits. Based on this, the decision of whether the self-vehicle SV should change lanes or keep lanes is given. Then the decision is passed to the decision-making evaluation submodule of the lane-changing performance baseline evaluation module, and the weights of the PV benefit function are passed to the weight self-learning submodule. The decision-making evaluation submodule determines whether the weights of the current PV benefit function have optimization space, and outputs the evaluation results to the weight self-learning submodule. The weight self-learning submodule receives the PV benefits from the decision-making submodule. The weights of the function, the evaluation results from the decision evaluation submodule and the driving style quantification data from the driving style quantification data temporary storage submodule. If the evaluation results from the decision evaluation submodule show that there is room for optimization in the weights of the PV benefit function, the weight self-learning submodule then guides the dynamic adjustment of the weights of the PV benefit function, and stores the new weights and the corresponding driving style quantification data into the driving style and dynamic weight storage submodule. Otherwise, the weight self-learning submodule does not adjust the weights of the PV benefit function, and directly passes the corresponding weights and the corresponding driving style quantification data to the driving style and dynamic weight storage submodule.

[0109] Specifically, at the current time step t, the lane change benefit function submodule calculates the SV benefit function result and PV benefit function result of the current time step t based on the kinematic information of the target lane rear vehicle at the current time step t obtained from the data set, the kinematic information of the ego vehicle, the kinematic information of the surrounding vehicles, and the current driving style quantification data obtained from the driving style quantification module, using the established ego vehicle SV benefit function and PV benefit function. The established ego vehicle SV benefit function and PV benefit function are as follows:

[0110] SV return function S priceSV for:

[0111]

[0112] In the above formula, K SV1 , K SV2 , K SV3 , Both are positive proportional coefficients, R lane is the lane-changing benefit of the SV when making a lane-changing decision;

[0113] V S (v SV ,v FV1 ) function represents the desired benefit of the ego vehicle SV to ensure its own driving speed by changing lanes, which is expressed as: Among them, v SV is the velocity of the ego vehicle SV at the current time step t, v FV1 is the velocity of FV1 at the current time step t;

[0114] V FV (v FV1 ,v FV2 ) function represents the benefit of the vehicle’s lane-changing action in improving its subsequent following speed, which can be expressed as: Among them, v FV1 is the velocity of FV1 at the current time step t, v FV2 is the velocity of FV2 at the current time step t;

[0115] S PV (S SV ,S PV ) function represents the safety benefit based on the longitudinal distance between the self-vehicle SV and the rear vehicle PV, which is expressed as:

[0116]

[0117] Among them, S SV is the longitudinal position of SV at the current time step t (with the vehicle's travel direction as the positive direction of the coordinate system), S PV is the longitudinal position of PV at the current time step t; S PV (S SV ,S PV ) is a piecewise function. When the longitudinal distance S between SV and PV is SV -S PV Exceeding the safety threshold S s , providing a constant positive return R sp ; When the longitudinal distance S between SV and PV SV -S PV At the safety threshold S s With the danger threshold S d When the spacing is between, it provides a positive return that gradually decreases as the spacing decreases. Until the risk threshold is reached, the return drops to 0; when the vertical distance S between SV and PV SV -S PV Below the danger threshold S d When the spd ;

[0118] Function representation The SV behavior decision constraint function based on the longitudinal distance between SV and FV1 and the longitudinal speed is expressed as: Among them, S FV1 is the longitudinal position of FV1 at the current time step t. When the longitudinal arrival time of SV and FV1 is When it is less than 3.5 seconds, Provides a large negative return R spn1 ; When the longitudinal arrival time of SV and FV1 When it is greater than or equal to 3.5 seconds, Provides a small positive return R sp1 ;

[0119] Function representation The SV behavior decision constraint function based on the longitudinal distance between SV and FV2 and the longitudinal speed is expressed as: Among them, S FV2 is the longitudinal position of FV2 at the current time step t. When the longitudinal arrival time of SV and FV2 is Less than the minimum lane-changing safety following time difference t sp2 =3.5+t lanec hour, Provides a large negative return R spn2 ; When the longitudinal arrival time of SV and FV2 Greater than or equal to the minimum lane-changing safety following time difference t sp2 hour,

[0120] Provides a small positive return R sp2 , where t lanec is the average lane changing time of the driver, t lanec The average lane changing time of drivers is obtained by counting the public data set;

[0121] When the SV does not change lanes and stays in the original lane, it obtains a constant small negative benefit R stay .

[0122] PV return function S pricePV for:

[0123] S pricePV =ω1·V PV (a PV ,v PV )+ω2·S Space (S PV ,S SV )

[0124] In the above formula, ω1 and ω2 represent the weights of the PV return function and are dynamically variable positive weight coefficients. PV (a PV ,v PV ) function represents the speed requirement benefit of PV, expressed as: Among them, v PV is the velocity of PV at the current time step t, a PV is the acceleration value of PV at the current time step t. When the action of PV is to avoid deceleration, a PV is a negative constant. When the PV action is not to avoid, a PV=0, τ0 is a positive constant; S Space (S PV ,S SV ) function represents the required return of PV, expressed as: Among them, h0 is a positive constant, S SV is the longitudinal position of SV at the current time step t, S PV is the longitudinal position of PV at the current time step t.

[0125] Specifically, the decision-making process of the decision judgment submodule is as follows: the decision judgment submodule obtains the SV benefit function result and the PV benefit function result of the current time step t from the lane change benefit function submodule, and first determines the lane change benefit R of the SV at this time. lane Is it greater than the lane keeping benefit R of staying in the original lane? stay , if the lane-changing benefit of SV is R lane No greater than the lane keeping benefit R of staying in the original lane stay , the decision-making submodule determines that the lane-changing prerequisite is not met at this time, and outputs the decision that the SV should keep the lane; if the lane-changing benefit R of SV at this time lane Greater than the lane keeping benefit R of staying in the original lane stay , the decision judgment submodule determines that the prerequisite for lane changing is met at this time; then the PV avoidance deceleration benefit and the PV non-avoidance benefit are calculated, and the benefit values ​​of the two are compared, and it is considered that PV will take an action with a larger benefit value: if the PV avoidance deceleration benefit is greater than or equal to the non-avoidance benefit, then it is judged that PV takes an avoidance deceleration action; if the PV avoidance deceleration benefit is less than the non-avoidance benefit, then it is judged that PV takes a non-avoidance action, and based on this, a decision is given as to whether the ego vehicle SV should change lanes or continue lane keeping at this time: if the PV action is judged to be avoidance deceleration, then the output decision of the decision judgment submodule at this time is lane changing; if the PV action is judged to be non-avoidance, then the output decision of the decision judgment submodule at this time is lane keeping, and then the decision judgment submodule passes the decision to the decision evaluation submodule of the lane changing performance baseline evaluation module, and passes the weights of the PV benefit function to the weight self-learning submodule.

[0126] like Figure 1 , 2As shown in Figure 5, the working process of the lane changing performance baseline evaluation module is as follows: first, the kinematic information of the target lane rear vehicle, the kinematic information of the vehicle itself, the kinematic information of the surrounding vehicles, and the driving style quantification data of the driving style quantification module extracted from the data set are used as the current state, and after data preprocessing, they are transmitted to the trained data-driven lane changing decision submodule. The trained data-driven lane changing decision submodule outputs the action of the current SV, i.e., lane changing or lane keeping, based on the current state transmitted, and then transmits the action to the decision evaluation submodule. The decision evaluation submodule determines the lane changing decision submodule based on the decision judgment submodule of the lane changing decision module. The decision of whether the SV should change lanes or keep lane, which is passed into the module, is compared with the output of the data-driven lane changing decision submodule, and the output decision result of the current data-driven lane changing decision submodule and the output decision result of the decision judgment submodule of the lane changing decision module are evaluated. Then the evaluation result is output to the baseline self-evaluator and the weight self-learning submodule of the lane changing decision module. The weight self-learning submodule dynamically adjusts the weights of the PV benefit function according to the evaluation result of the decision evaluation submodule. The baseline self-evaluator decides whether the current weight self-evolution process can be terminated according to the evaluation result of the decision evaluation submodule.

[0127] Specifically, the data preprocessing process before being passed into the trained data-driven lane-changing decision submodule is as follows:

[0128] The kinematic information of the target lane vehicle behind, the vehicle kinematic information, the surrounding vehicle kinematic information, and the driving style quantification data X obtained from the driving style quantification module are d Composition state dataset S D :

[0129]

[0130] Among them, i represents that the state is the i-th time series extracted from the data set, 0≤i≤N E ; For the i-th time sequence state sequence S D(i) , expressed as:

[0131]

[0132] Where t represents the current state data as the state data of the tth sampling point, 0≤t≤T E , the i-th time series S D(i,t) The state data at the tth sampling point is expressed as:

[0133] S D(i,t) =[E PV(i,t) ,E SV(i,t) ,E FV1(i,t) ,E FV2(i,t) ,X d(i) ]

[0134] Among them, E PV(i,t) represents the kinematic information of the vehicle PV behind the target lane at the tth sampling point, E SV(i,t) represents the kinematic information of the ego vehicle SV at the tth sampling point, E FV1(i,t) represents the kinematic information of the front vehicle FV1 in the lane before the ego vehicle changes lanes at the tth sampling point, E FV2(i,t) represents the kinematic information of the front vehicle FV2 in the target lane at the tth sampling point, X d(i) Represents the driving style quantification data of the PV of the i-th time series.

[0135] Specifically, the training process of the data-driven lane-changing decision submodule is as follows:

[0136] First, from the state data set S D Take any time sequence S from D(i) , and then the state data of each sampling point of the time series state sequence is sequentially transmitted to the data-driven lane-changing decision submodule in time sequence; at the tth time step, S D(i,t) The deep Q network in the data-driven lane-changing decision submodule calculates the value of the current action state. And make an action t , where action a t is one of two discrete actions: lane keeping or lane changing. is the model parameter of the deep Q network, and then comes the t+1th time step, S D(i,t+1) The deep Q network in the data-driven lane-changing decision submodule obtains the reward function R S (t) Define the true reward r t , based on maximizing the action state value at this time The principle of estimating action a t+1 = argmaxQ(s t+1 ,a,ω), calculate the timing target Where 0<γ<1, which is the minimum loss Loss: Perform gradient descent to update model parameters: Where η is a positive constant;

[0137] Repeat the above steps until the model parameters converge, and the data-driven lane-changing decision submodule is trained. D(i,t) The data-driven lane-changing decision submodule is input, and the data-driven lane-changing decision submodule outputs the decision at time step t.

[0138] Furthermore, the reward function R of the deep Q network of the data-driven lane-changing decision submodule is S(t) is defined as follows:

[0139]

[0140] Among them, μ1, μ2, μ3 are constant coefficients, v SV (t) indicates the SV(i,t) The obtained SV longitudinal velocity information, v FV1 (t) indicates the FV1(i,t) The longitudinal velocity information of FV1 is obtained, v FV2 (t) indicates the FV2(i,t) The longitudinal velocity information of FV2 is obtained, S SV (t) indicates the SV(i,t) The longitudinal position information of the SV obtained, S PV (t) indicates the PV(i,t) The longitudinal position information of the PV obtained, S d is the danger threshold.

[0141] Specifically, the evaluation process of the decision evaluation submodule is as follows:

[0142] The decision evaluation submodule receives the decision action A1 (i.e., SV lane change or lane keeping) from the decision judgment submodule of the lane change decision module and the decision action A2 (i.e., SV lane change or lane keeping) from the data-driven lane change decision submodule. Based on the output values ​​of these two actions, the decision evaluation submodule determines the quality of the action and takes action according to the following judgment rules:

[0143] (1) If A1 = A2 = lane keeping, then at this moment, since lane changing behavior has not occurred, the decision evaluation submodule only outputs decision-consistent information to the baseline self-evaluator.

[0144] (2) If A1 = lane keeping and A2 = lane changing, then the decision evaluation submodule uses the cubic polynomial method to keep the speed at the start and end of the lane change consistent, and the lane change duration is t lanec Plan the lane change route and try to change lanes. Cubic polynomial path planning is a common method for path planning, and its specific details are not repeated here. During the lane change attempt, the driving trajectories of the surrounding vehicles and PV are updated based on the kinematic information obtained from the data set. The driving trajectory of the SV is no longer updated based on the kinematic information obtained from the data set, but is updated based on the planned lane change route: If the longitudinal distance between the SV and the PV is lower than the safety threshold S during the lane change attempt, the SV will be automatically updated. d , it is considered that lane changing is inferior to lane keeping at this moment, so the decision evaluation submodule only outputs the information that is better than the lane changing decision module to the baseline self-evaluator; if the longitudinal distance between SV and PV is always higher than the safety threshold S during the lane changing attempt, d, it is considered that lane changing is better than lane keeping at this moment, so the decision evaluation submodule outputs the information that lane changing is currently possible to the weight self-learning submodule of the lane changing decision module, and outputs the information of decision inconsistency to the baseline self-evaluator.

[0145] (3) If A1 = lane change, A2 = lane change, then at this moment, since the actions are consistent, the decision evaluation submodule only outputs the information of consistent decisions to the baseline self-evaluator.

[0146] (4) If A1 = lane change, A2 = lane keeping, then at this moment the decision evaluation submodule plans the lane change route using a cubic polynomial and attempts a lane change; during the lane change attempt, the driving trajectories of the surrounding vehicles and PV are updated based on the kinematic information obtained from the data set, while the driving trajectory of the SV is no longer updated based on the kinematic information obtained from the data set, but is updated based on the planned lane change route: If the longitudinal distance between the SV and PV is lower than the safety threshold S during the lane change attempt, d , it is considered that lane changing at this moment is inferior to lane keeping. Therefore, the decision evaluation submodule outputs information about the risk of the current lane change to the weight self-learning submodule of the lane change decision module, and outputs information about inconsistent decisions to the baseline self-evaluator. If the longitudinal distance between SV and PV is always higher than the safety threshold S during the lane change attempt, d , it is considered that lane changing is better than lane keeping at this moment, so the decision evaluation submodule only outputs the better information of the lane changing decision module to the baseline self-evaluator.

[0147] Specifically, the operation process of the baseline self-assessor is as follows:

[0148] The baseline self-evaluator receives the information output by the decision evaluation submodule and counts the number of information h1 that is better than the lane change decision module, the number of information h2 that is inconsistent with the decision, and the number of information h3 that is consistent with the decision every M (M is a positive constant) data. and It is considered that the decision of the decision judgment submodule is significantly better than the decision of the data-driven lane change decision submodule at this time, and the weight self-evolution process is terminated, otherwise the weight self-evolution process continues; β0 and β1 are positive constants greater than zero.

[0149] Specifically, the operation process of the weight self-learning submodule is as follows:

[0150] The weight self-learning submodule receives the weights of the PV benefit function of the decision judgment submodule and the information output by the decision evaluation submodule:

[0151] (1) When receiving information from the decision evaluation submodule that the current lane change is risky, it is considered that the weight of the speed benefit term in the PV benefit function is underestimated at this time. Therefore, the weight self-learning submodule adjusts the PV benefit function S according to the following formula: pricePV The weights ω1 and ω2 are the new weights ω1' and ω'2:

[0152] 1-ω1'→ω'2, where α is a positive constant coefficient.

[0153] (2) When receiving the information that the lane can be changed from the decision evaluation submodule, it is considered that the weight of the distance benefit term in the PV benefit function is underestimated at this time. Therefore, the weight self-learning submodule adjusts the PV benefit function S according to the following formula: pricePV The weights ω1 and ω2 are the new weights ω1' and ω'2:

[0154] 1-ω'2→ω1', where α is a positive constant coefficient.

[0155] (3) When no information is received from the decision evaluation submodule, the weight self-learning submodule does not adjust the weights of the PV benefit function.

[0156] Subsequently, the weight self-learning submodule transfers the latest weights of the PV benefit function and its corresponding driving style quantification data to the driving style and dynamic weight storage submodule for storage.

[0157] Specifically, the operation process of the driving style and dynamic weight storage submodule is as follows:

[0158] First, the driving style and dynamic weight storage submodule receives the weights of the PV benefit function and its corresponding driving style quantitative data newly transmitted from the weight self-learning submodule; then, the driving style and dynamic weight storage submodule queries whether there is a driving style quantitative data record in the current storage that is consistent with the newly transmitted driving style quantitative data, and performs the following operations:

[0159] (1) If there is a driving style quantization data record that is consistent with the newly-input driving style quantization data, the driving style and dynamic weight storage submodule updates the original record, and overwrites the weight records of the original PV benefit function with the weights of the newly-input PV benefit function.

[0160] (2) If there is no driving style quantification data record that is consistent with the newly-input driving style quantification data, the driving style and dynamic weight storage submodule stores the weights of the newly-input PV benefit function and its corresponding driving style quantification data as a record in the storage.

[0161] The overall process of weight self-evolution learning of a weight self-evolution intelligent vehicle personalized lane change decision system provided by the present invention is as follows:

[0162] Step 1: extract the kinematic information of the target lane rear vehicle, the ego vehicle kinematic information and the surrounding vehicle kinematic information from the data set, and then send the above information to the driving style quantification module, the lane change decision module and the lane change performance baseline evaluation module; first, randomly extract a driving information from the data set, and a driving information includes the kinematic information of the target lane rear vehicle, the ego vehicle kinematic information and the surrounding vehicle kinematic information at each time step of a lane change process; at time step t, the information input to the driving style quantification module is the kinematic information of the target lane rear vehicle of the driving information during the period from time step t-T+1 to time step t, a total of T time; at time step t, the information input to the lane change decision module and the lane change performance baseline evaluation module is the kinematic information of the target lane rear vehicle, the ego vehicle kinematic information and the surrounding vehicle kinematic information of the driving information at time step t.

[0163] Step 2: The driving style quantification module first transfers the kinematic information of the vehicle behind the target lane to the kinematic feature extraction submodule. The kinematic feature extraction submodule performs data preprocessing and data feature extraction on the kinematic information, and then transfers the processed information to the driving style evaluation submodule. The trained Gaussian mixture model in the driving style evaluation submodule receives the above information and gives the quantified probability of each driving style of the vehicle behind the target lane based on the information obtained, including: the probability of cautious type x d1 , the general probability x d2 , the probability of radical type x d3 , and the probability x d1 、x d2 、x d3 Output to the driving style quantization data temporary storage submodule, which organizes the data into a vector X with one row and three columns d =[x d1 ,x d2 ,x d3 ], i.e., driving style quantization data, and is stored in a data register; when new driving style quantization data is transmitted to the driving style quantization data temporary storage submodule, the previous data stored in the data register will be overwritten by the new quantization data.

[0164] Step three: the lane changing decision module calculates the benefit functions of SV and PV through the lane changing benefit function submodule based on the kinematic information of the vehicle behind the target lane, the kinematic information of the vehicle itself and the kinematic information of the surrounding vehicles and the driving style quantification data of the driving style quantification module, and outputs the calculation results and the weights of the PV benefit function to the decision judgment submodule; the decision judgment submodule outputs the decision judgment results to the decision evaluation submodule of the lane changing performance baseline evaluation module, and outputs the weights of the PV benefit function to the weight self-learning submodule.

[0165] Step 4: The data-driven lane-changing decision submodule of the lane-changing performance baseline evaluation module outputs the decision result to the decision evaluation submodule based on the kinematic information of the vehicle behind the target lane, the kinematic information of the vehicle itself, the kinematic information of the surrounding vehicles and the driving style quantification data of the driving style quantification module.

[0166] Step 5: The decision evaluation submodule evaluates which one is better, the output decision result of the current data-driven lane changing decision submodule or the output decision result of the decision judgment submodule of the lane changing decision module, and then outputs the evaluation result to the weight self-learning submodule and the baseline self-evaluator of the lane changing decision module.

[0167] Step six: the weight self-learning submodule dynamically adjusts the weights of the PV benefit function according to the evaluation results of the decision evaluation submodule, and transmits the latest weights of the PV benefit function and the corresponding driving style quantification data obtained from the driving style quantification data temporary storage submodule to the driving style and dynamic weight storage submodule of the lane change decision module. The driving style and dynamic weight storage submodule stores the transmitted information.

[0168] In step seven, the baseline self-evaluator decides whether the current weight self-evolution process can be terminated based on the evaluation results of the decision evaluation submodule. If the weight has not been adjusted to the optimal state, steps one to seven are repeated; if the weight has been adjusted to the optimal state, the weight self-evolution process is terminated.

[0169] See also Figure 2 and Figure 6 , when the baseline self-evaluator determines that the weight self-learning process can be terminated, it is considered that the weight has been adjusted to the optimal state. At this time, the lane change decision system is trained and has the ability to make lane change decisions in a real traffic environment. The following is a practical application process:

[0170] Step 1: The ego vehicle SV obtains the kinematic information of the vehicle behind the current target lane, the ego vehicle kinematic information, and the kinematic information of the surrounding vehicles through the laser radar, camera and other sensing devices and the ego vehicle kinematic monitoring equipment, and then sends the above information to the driving style quantification module, the lane change decision module and the lane change performance baseline evaluation module. At time step t, the information input to the driving style quantification module is the kinematic information of the vehicle behind the target lane from time step t-T+1 to time step t, a total of T time; at time step t, the information input to the lane change decision module and the lane change performance baseline evaluation module is the kinematic information of the vehicle behind the target lane, the ego vehicle kinematic information, and the kinematic information of the surrounding vehicles at time step t.

[0171] Step 2: The driving style quantification module gives the driving style quantification result X based on the kinematic information of the vehicle behind the target lane. di , output to the lane-changing decision module.

[0172] Step 3: The driving style and dynamic weight storage submodule of the lane change decision module quantifies the result X according to the driving style di , find the weight stored during evolution and the driving style quantization result X di The driving style quantification result X with the smallest Euclidean distance dj , X dj The corresponding dynamic weights and the kinematic information of the vehicle behind the target lane, the vehicle kinematic information and the surrounding vehicle kinematic information are output to the lane change benefit function submodule.

[0173] Step 4: The lane-changing benefit function submodule calculates the benefit functions of SV and PV, and outputs the calculation results to the decision-making submodule.

[0174] In step 5, the decision judgment submodule outputs the final decision action, that is, the decision to change lanes or lane keeping. If the final decision action is to change lanes, the decision process ends and the lane change signal is handed over to the vehicle controller to execute the lane change action; if the final decision action is to keep lanes, steps 1 to 5 are repeated.

Claims

1. A weighted self-evolving intelligent vehicle personalized lane-changing decision system, characterized by: It includes driving style quantification module, lane change decision module, and lane change performance baseline assessment module; The driving style quantification module includes a kinematic feature extraction submodule, a driving style evaluation submodule and a driving style quantification data temporary storage submodule; The lane-changing decision module includes: a lane-changing benefit function submodule, a decision judgment submodule, a weight self-learning submodule, and a driving style and dynamic weight storage submodule; the lane-changing performance baseline evaluation module includes a data-driven lane-changing decision submodule, a decision evaluation submodule, and a baseline self-evaluator; The working process of the driving style quantification module is as follows: first, the kinematic information of the vehicle behind the target lane is transmitted to the kinematic feature extraction submodule, the kinematic feature extraction submodule performs data preprocessing and data feature extraction on the transmitted kinematic information, and then transmits the processed information to the driving style evaluation submodule, the trained Gaussian mixture model in the driving style evaluation submodule receives the above information, and gives the quantified probability of each driving style of the vehicle behind the target lane according to the obtained information, including: the probability of cautious type x d1 , the general probability x d2 , the probability of radical type x d3 , and the probability x d1 、x d2 、x d3 Output to the driving style quantization data temporary storage submodule, which organizes the data into a vector X with one row and three columns d =[x d1 ,x d2 ,x d3 ], i.e., driving style quantization data, and is stored in a data register; when new driving style quantization data is transmitted to the driving style quantization data temporary storage submodule, the previous data stored in the data register will be overwritten by the new quantization data; The working process of the lane change decision module is as follows: the kinematic information of the vehicle behind the target lane, the kinematic information of the vehicle itself, the kinematic information of the surrounding vehicles, and the driving style quantification data X of the driving style quantification module are combined. d =[x d1 ,x d2 ,x d3 ] is passed to the lane-changing benefit function submodule, which is based on the established benefit function of the ego vehicle SV and the benefit function of the vehicle behind the target lane PV; first query whether there is a dynamic weight record of the same driving style quantified data in the driving style and dynamic weight storage submodule, if there is a record, then call the dynamic weight record as the weights ω1 and ω2 of the PV benefit function; if there is no record, then according to the driving style quantified data X d =[x d1 ,x d2 ,x d3 ]Preliminary adjustment of the weights of the PV return function: Then, the SV benefit function result is calculated based on the established benefit function of the ego vehicle SV, and the PV benefit function result is calculated based on the established benefit function of the vehicle behind the target lane PV; the calculated SV benefit function result, PV benefit function result and each weight of the PV benefit function are transmitted to the decision judgment submodule, and the decision judgment submodule first determines the lane change benefit R of the SV at this time lane Is it greater than the lane keeping benefit R of staying in the original lane? stay , if the lane-changing benefit of SV is R lane No greater than the lane keeping benefit R of staying in the original lane stay , then the decision of the vehicle SV to keep the lane is given; if the lane change benefit R lane Greater than the lane keeping benefit R of staying in the original lane stay , the decision-making submodule is based on the game idea that PV expects to obtain the maximum benefit. It defaults that PV will take the action of evading and slowing down or continuing to drive without evading according to the principle of maximizing its own benefits. Based on this, the decision of whether the self-vehicle SV should change lanes or keep lanes is given. Then the decision is passed to the decision-making evaluation submodule of the lane-changing performance baseline evaluation module, and the weights of the PV benefit function are passed to the weight self-learning submodule. The decision-making evaluation submodule determines whether the weights of the current PV benefit function have optimization space, and outputs the evaluation results to the weight self-learning submodule. The weight self-learning submodule receives the PV benefits from the decision-making submodule. The weights of the function, the evaluation results from the decision evaluation submodule and the driving style quantitative data from the driving style quantitative data temporary storage submodule. If the evaluation results from the decision evaluation submodule show that there is room for optimization in the weights of the PV benefit function, the weight self-learning submodule then guides the dynamic adjustment of the weights of the PV benefit function, and stores the new weights and the corresponding driving style quantitative data in the driving style and dynamic weight storage submodule. Otherwise, the weight self-learning submodule does not adjust the weights of the PV benefit function, and directly transfers the corresponding weights and the corresponding driving style quantitative data to the driving style and dynamic weight storage submodule. The working process of the lane changing performance baseline evaluation module is as follows: first, the kinematic information of the target lane rear vehicle, the kinematic information of the vehicle, the kinematic information of the surrounding vehicles, and the driving style quantification data of the driving style quantification module extracted from the data set are used as the current state, and after data preprocessing, they are transmitted to the trained data-driven lane changing decision submodule. The trained data-driven lane changing decision submodule outputs the action of the current SV, that is, lane changing or lane keeping, based on the current state transmitted. Then, the action is transmitted to the decision evaluation submodule. The decision evaluation submodule compares the decision of the SV that should currently change lanes or keep lanes transmitted by the decision judgment submodule of the lane changing decision module with the output of the data-driven lane changing decision submodule, and evaluates which of the output decision results of the current data-driven lane changing decision submodule and the output decision results of the decision judgment submodule of the lane changing decision module is better. Then, the evaluation results are output to the baseline self-evaluator and the weight self-learning submodule of the lane changing decision module. The weight self-learning submodule dynamically adjusts the weights of the PV benefit function according to the evaluation results of the decision evaluation submodule. The baseline self-evaluator determines whether the current weight self-evolution process can be terminated according to the evaluation results of the decision evaluation submodule.

2. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: The data feature extraction function of the kinematic feature extraction submodule is to obtain the relative speed, relative acceleration and jerk information of the vehicle behind the target lane relative to the vehicle within the time T, calculate the average relative speed, average relative acceleration and average jerk, and pass the calculated average relative speed, average relative acceleration and average jerk information to the driving style evaluation submodule.

3. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: The training process of the Gaussian mixture model is as follows: Step 1: Preprocessing and feature extraction: Obtain the relative speed, relative acceleration and jerk information of the target lane rear vehicle with N lane-changing behaviors in the data set within T time before the ego vehicle changes lanes, and calculate the average relative speed of the target lane rear vehicle in the i-th training data. Average relative acceleration and average jerk Recorded as All N qualified training data are expressed as: P = [P1, P2, ..., P n ]; Step 2: Parameter training of Gaussian mixture model: All N qualified training data are passed into the Gaussian mixture model. The number of sub-Gaussian models K in the Gaussian mixture model is set to 3. The maximum expectation algorithm is used to solve the parameters of the Gaussian mixture model. It is mainly divided into two steps: The first step, E-step, calculates the Q function according to the following formula, which represents the probability that the i-th data belongs to the k-th sub-Gaussian distribution: Among them, φ k represents the probability of the kth sub-Gaussian distribution, x (i) represents the i-th data, μ k represents the mean of the kth sub-Gaussian distribution, σ k Represents the standard deviation of the k-th sub-Gaussian distribution, and φ is randomly given in the first iteration k , μ k , σ k The value of The second step, M-step, updates the model parameters according to the following formula: Repeat the iteration of E-step and M-step several times until the model converges; After the training is completed, the Gaussian mixture model evaluates the probability x that the driving style of the following vehicle is cautious by inputting the average relative speed, average relative acceleration, and average jerk information of the following vehicle within the time T before the self-vehicle changes lanes. d1 , the general probability x d2 , the probability of radical type x d3 , and output the above probability as the output of the driving style evaluation submodule.

4. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: At the current time step t, the lane change benefit function submodule calculates the SV benefit function result and PV benefit function result of the current time step t based on the kinematic information of the target lane rear vehicle, the ego vehicle kinematic information, the surrounding vehicle kinematic information and the current driving style quantification data obtained from the driving style quantification module, using the established ego vehicle SV benefit function and PV benefit function. The established ego vehicle SV benefit function and PV benefit function are as follows: SV return function S priceSV for: In the above formula, K SV1 , K SV2 , K SV3 , K p1 , K p2 Both are positive proportional coefficients, R lane is the lane-changing benefit of SV when making lane-changing decision; V S (v SV ,v FV1 ) function represents the desired benefit of the ego vehicle SV to ensure its own driving speed by changing lanes, which is expressed as: Among them, v SV is the velocity of the ego vehicle SV at the current time step t, v FV1 is the speed of the front vehicle FV1 in the lane where the ego vehicle was before changing lanes at the current time step t; V FV (v FV1 ,v FV2 ) function represents the benefit of the vehicle’s lane-changing action in improving its subsequent following speed, which can be expressed as: Among them, v FV1 is the velocity of FV1 at the current time step t, v FV2 is the speed of the front vehicle FV2 in the target lane at the current time step t; S PV (S SV ,S PV ) function represents the safety benefit based on the longitudinal distance between the self-vehicle SV and the rear vehicle PV, which is expressed as: Among them, S SV is the longitudinal position of SV at the current time step t, S PV is the longitudinal position of PV at the current time step t; S PV (S SV ,S PV ) is a piecewise function. When the longitudinal distance S between SV and PV is SV -S PV Exceeding the safety threshold S s Provides a constant positive return R sp ; When the longitudinal distance S between SV and PV SV -S PV At the safety threshold S s With the danger threshold S d When the spacing is between, it provides a positive return that gradually decreases as the spacing decreases. Until the risk threshold is reached, the return drops to 0; when the vertical distance S between SV and PV SV -S PV Below the danger threshold S d When the spd ; S p1 (S PV1 ,S SV ) function represents the SV behavior decision constraint function based on the longitudinal distance between SV and FV1 and the longitudinal speed, which is expressed as: Among them, S FV1 is the longitudinal position of FV1 at the current time step t. When the longitudinal arrival time of SV and FV1 is When it is less than 3.5 seconds, S p1 (S FV1 ,S SV ) provides a large negative return R spn1 ; When the longitudinal arrival time of SV and FV1 When it is greater than or equal to 3.5 seconds, S p1 (S FV1 ,S SV ) provides a small positive return R sp1 ; S p2 (S FV2 ,S SV ) function represents the SV behavior decision constraint function based on the longitudinal distance between SV and FV2 and the longitudinal speed, which is expressed as: Among them, S FV2 is the longitudinal position of FV2 at the current time step t. When the longitudinal arrival time of SV and FV2 is Less than the minimum lane-changing safety following time difference t sp2 =3.5+t lanec When S p2 (S FV2 ,S SV ) provides a large negative return R spn2 ; When the longitudinal arrival time of SV and FV2 Greater than or equal to the minimum lane-changing safety following time difference t sp2 hour, Provides a small positive return R sp2 , where t lanec is the average lane changing time of the driver, t lanec The average lane changing time of drivers is obtained by counting the public data set; When the SV does not change lanes and stays in the original lane, it obtains a constant small negative benefit R stay ; PV return function S pricePV for: S pricePV =ω1·V PV (a PV ,v PV )+ω2·S Space (S PV ,S SV ) In the above formula, ω1 and ω2 represent the weights of the PV return function and are dynamically variable positive weight coefficients. PV (a PV ,v PV ) function represents the speed requirement benefit of PV, expressed as: Among them, v PV is the velocity of PV at the current time step t, a PV is the acceleration value of PV at the current time step t. When the action of PV is to avoid deceleration, a PV is a negative constant. When the PV action is not to avoid, a PV =0, τ0 is a positive constant; S Space (S PV ,S SV ) function represents the required return of PV, expressed as: Among them, h0 is a positive constant, S SV is the longitudinal position of SV at the current time step t, S PV is the longitudinal position of PV at the current time step t.

5. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: The decision-making process of the decision judgment submodule is as follows: the decision judgment submodule obtains the SV benefit function result and PV benefit function result of the current time step t from the lane change benefit function submodule, and first determines the lane change benefit R of the SV at this time. lane Is it greater than the lane keeping benefit R of staying in the original lane? stay , if the lane-changing benefit of SV is R lane No greater than the lane keeping benefit R of staying in the original lane stay , the decision-making submodule determines that the lane-changing prerequisite is not met at this time, and outputs the decision that the SV should keep the lane; if the lane-changing benefit R of SV at this time lane Greater than the lane keeping benefit R of staying in the original lane stay , the decision judgment submodule determines that the prerequisite for lane changing is met at this time; then the PV avoidance deceleration benefit and the PV non-avoidance benefit are calculated, and the benefit values ​​of the two are compared, and it is considered that PV will take an action with a larger benefit value: if the PV avoidance deceleration benefit is greater than or equal to the non-avoidance benefit, then it is judged that PV takes an avoidance deceleration action; if the PV avoidance deceleration benefit is less than the non-avoidance benefit, then it is judged that PV takes a non-avoidance action, and based on this, a decision is given as to whether the ego vehicle SV should change lanes or continue lane keeping at this time: if the PV action is judged to be avoidance deceleration, then the output decision of the decision judgment submodule at this time is lane changing; if the PV action is judged to be non-avoidance, then the output decision of the decision judgment submodule at this time is lane keeping, and then the decision judgment submodule passes the decision to the decision evaluation submodule of the lane changing performance baseline evaluation module, and passes the weights of the PV benefit function to the weight self-learning submodule.

6. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: The process of data preprocessing before passing it into the trained data-driven lane-changing decision submodule is as follows: The kinematic information of the target lane vehicle behind, the vehicle kinematic information, the surrounding vehicle kinematic information, and the driving style quantification data X obtained from the driving style quantification module are d Composition state dataset S D : Among them, i represents that the state is the i-th time series extracted from the data set, 0≤i≤N E ; For the i-th time sequence state sequence S D(i) , expressed as: Where t represents the current state data as the state data of the tth sampling point, 0≤t≤T E , the i-th time series S D(i,t) The state data at the tth sampling point is expressed as: S D(i,t) =[And PV(i,t) ,AND SV(i,t) ,AND FV1(i,t) ,AND FV2(i,t) ,X d(i) ] Among them, E PV(i,t) represents the kinematic information of the vehicle PV behind the target lane at the tth sampling point, E SV(i,t) represents the kinematic information of the ego vehicle SV at the tth sampling point, E FV1(i,t) represents the kinematic information of the front vehicle FV1 in the lane before the ego vehicle changes lanes at the tth sampling point, E FV2(i,t) represents the kinematic information of the front vehicle FV2 in the target lane at the tth sampling point, X d(i) Represents the driving style quantification data of the PV of the i-th time series.

7. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: The training process of the data-driven lane-changing decision submodule is as follows: First, from the state data set S D Take any time sequence S from D(i) , and then the state data of each sampling point of the time series state sequence is sequentially transmitted to the data-driven lane-changing decision submodule in time sequence; at the tth time step, S D(i,t) The deep Q network in the data-driven lane-changing decision submodule calculates the value of the current action state. And make an action t , where action a t is one of two discrete actions: lane keeping or lane changing. is the model parameter of the deep Q network, and then comes the t+1th time step, S D(i,t+1) The deep Q network in the data-driven lane-changing decision submodule obtains the reward function R S (t) Define the true reward r t , based on maximizing the action state value at this time The principle of estimating action a t+1 = argmaxQ(s t+1 ,a,ω), calculate the timing target Where 0<γ<1, which is the minimum loss Loss: Perform gradient descent to update model parameters: Where η is a positive constant; Repeat the above steps until the model parameters converge, and the data-driven lane-changing decision submodule is trained. After the training is completed, the S D(i,t) The data-driven lane-changing decision submodule is input, and the data-driven lane-changing decision submodule outputs the decision at time step t.

8. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 7, characterized in that: The reward function R of the deep Q network is S (t) is defined as follows: Among them, μ1, μ2, μ3 are constant coefficients, v SV (t) indicates the SV(i,t) The obtained SV longitudinal velocity information, v FV1 (t) indicates the FV1(i,t) The longitudinal speed information of the front vehicle FV1 in the lane where the vehicle is located before changing lanes is obtained, v FV2 (t) indicates the FV2(i,t) The longitudinal speed information of the front vehicle FV2 in the target lane is obtained, S SV (t) indicates the SV(i,t) The longitudinal position information of the SV obtained, S PV (t) indicates the PV(i,t) The longitudinal position information of the PV obtained, S d is the danger threshold.

9. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: The evaluation process of the decision evaluation submodule is as follows: The decision evaluation submodule receives the decision action A1 from the decision judgment submodule of the lane change decision module and the decision action A2 from the data-driven lane change decision submodule, and determines the quality of the action and takes action according to the output values ​​of the two actions according to the following judgment rules: (1) If A1 = A2 = lane keeping, then since lane changing has not occurred, the decision evaluation submodule only outputs decision-consistent information to the baseline self-evaluator; (2) If A1 = lane keeping and A2 = lane changing, then the decision evaluation submodule uses the cubic polynomial method to keep the speed at the start and end of the lane change consistent, and the lane change duration is t lanec Plan a lane-changing route and attempt to change lanes; during the lane-changing attempt, the trajectories of the surrounding vehicles and PV are updated based on the kinematic information obtained from the data set, and the SV's trajectory is updated based on the planned lane-changing route: If the longitudinal distance between SV and PV is lower than the safety threshold S during the lane-changing attempt, d , it is considered that lane changing is inferior to lane keeping at this moment, so the decision evaluation submodule only outputs the information that the lane changing decision module is better to the baseline self-evaluator; If the longitudinal distance between SV and PV is always higher than the safety threshold S during the lane change attempt, d , it is considered that lane changing is better than lane keeping at this moment, so the decision evaluation submodule outputs the information that lane changing is possible at present to the weight self-learning submodule of the lane changing decision module, and outputs the information of decision inconsistency to the baseline self-evaluator; (3) If A1 = lane change, A2 = lane change, then at this moment, since the actions are consistent, the decision evaluation submodule only outputs the decision-consistent information to the baseline self-evaluator; (4) If A1 = lane change, A2 = lane keeping, then the decision evaluation submodule uses a cubic polynomial to plan the lane change route and attempts to change lanes. During the lane change attempt, the trajectories of the surrounding vehicles and PV are updated based on the kinematic information obtained from the data set, and the SV's trajectory is updated based on the planned lane change route. If the longitudinal distance between SV and PV is lower than the safety threshold S during the lane change attempt, d , it is considered that lane changing at this moment is inferior to lane keeping, so the decision evaluation submodule outputs information that there is a risk in the current lane change to the weight self-learning submodule of the lane change decision module, and outputs information about inconsistent decisions to the baseline self-evaluator; If the longitudinal distance between SV and PV is always higher than the safety threshold S during the lane change attempt, d , it is considered that lane changing is better than lane keeping at this moment, so the decision evaluation submodule only outputs the better information of the lane changing decision module to the baseline self-evaluator; The operation process of the baseline self-assessor is as follows: The baseline self-evaluator receives the information output by the decision evaluation submodule and counts the number of information h1 that is better than the lane change decision module, the number of information h2 that is inconsistent with the decision, and the number of information h3 that is consistent with the decision. and It is considered that the decision of the decision judgment submodule is significantly better than the decision of the data-driven lane change decision submodule at this time, and the weight self-evolution process is terminated, otherwise the weight self-evolution process continues; β0 and β1 are positive constants greater than zero.

10. The weight self-evolving intelligent vehicle personalized lane-changing decision system according to claim 1, characterized in that: The operation process of the weight self-learning submodule is as follows: The weight self-learning submodule receives the weights of the PV benefit function of the decision judgment submodule and the information output by the decision evaluation submodule: (1) When receiving information from the decision evaluation submodule that the current lane change is risky, it is considered that the weight of the speed benefit term in the PV benefit function is underestimated at this time. Therefore, the weight self-learning submodule adjusts the PV benefit function S according to the following formula: pricePV The weights ω1 and ω2 are the new weights ω'1 and ω'2: 1-ω'1→ω'2, where α is a positive constant coefficient; (2) When receiving the information that the lane can be changed from the decision evaluation submodule, it is considered that the weight of the distance benefit term in the PV benefit function is underestimated at this time. Therefore, the weight self-learning submodule adjusts the PV benefit function S according to the following formula: pricePV The weights ω1 and ω2 are the new weights ω'1 and ω'2: 1-ω'2→ω'1, where α is a positive constant coefficient; (3) When no information is received from the decision evaluation submodule, the weight self-learning submodule does not adjust the weights of the PV benefit function; Subsequently, the weight self-learning submodule transfers the latest weights of the PV benefit function and its corresponding driving style quantitative data to the driving style and dynamic weight storage submodule for storage; The operation process of the driving style and dynamic weight storage submodule is as follows: First, the driving style and dynamic weight storage submodule receives the weights of the PV benefit function and its corresponding driving style quantitative data newly transmitted from the weight self-learning submodule; then, the driving style and dynamic weight storage submodule queries whether there is a driving style quantitative data record in the current storage that is consistent with the newly transmitted driving style quantitative data, and performs the following operations: (1) If there is a driving style quantization data record that is consistent with the newly-input driving style quantization data, the driving style and dynamic weight storage submodule updates the original record, and the weights of the newly-input PV benefit function are used to overwrite the weight records of the original PV benefit function; (2) If there is no driving style quantification data record that is consistent with the newly-input driving style quantification data, the driving style and dynamic weight storage submodule stores the weights of the newly-input PV benefit function and its corresponding driving style quantification data as a record in the storage.

Images