CN117877245A - Novel heterogeneous mixed traffic flow model grading evaluation and construction method - Google Patents

Abstract

The invention discloses a novel grading evaluation and construction method of a heterogeneous mixed traffic flow model, which comprises the following steps: step one, establishing an automatic driving automobile function grading evaluation index; secondly, designing a decision control algorithm for the automatic driving function corresponding to each stage according to the grading index; and thirdly, constructing a mixed traffic flow model. The beneficial effects are that: the grading index which better accords with the traffic flow characteristics is established, and meanwhile, the uniform distribution and the dispersion of different functions, namely different grades of automatic driving vehicles in a given road are realized, so that a heterogeneous mixed traffic flow simulation model is established. The method can be applied to heterogeneous mixed traffic flow models with high confidence coefficient for intelligent automobile simulation research, can be popularized and applied to research in the fields of intelligent automobile decision planning, control, testing and the like, and has wide application prospects.

Description

Novel heterogeneous mixed traffic flow model grading evaluation and construction method

Technical Field

The invention relates to a model grading evaluation and construction method, in particular to a novel heterogeneous mixed traffic flow model grading evaluation and construction method.

Background

At present, with the development of intelligent internet-connected automobiles and automatic driving, the ratio of vehicles with different automatic driving functions in a road is gradually increased, so that the interaction between the behaviors of the vehicles is more complex and changeable.

Most of the mixed traffic flow simulation models in the current stage are built based on a single model, the behaviors and influences of different automatic driving function vehicles in traffic flows are difficult to reflect, a unified framework or an evaluation method is lacked to evaluate and develop heterogeneous mixed traffic flow models, the heterogeneous mixed traffic flow models refer to traffic flow models with interaction and dynamic change among different types of vehicles, for example, CN112329248A discloses a road mixed traffic flow simulation system based on a multi-agent system, but coordination and conflict among the automatic driving vehicles with different functions are not considered in the method, and uniform distribution and dispersion of the models are not performed.

The method for classifying and evaluating the heterogeneous mixed traffic flow is very significant in designing corresponding traffic flow models.

Disclosure of Invention

The invention aims to solve the problems that the existing traffic flow simulation model cannot truly reflect traffic situations of different functions of automatic driving vehicles in a road and is difficult to meet high-confidence simulation requirements of heterogeneous mixed traffic flows, and the like, and provides a novel heterogeneous mixed traffic flow model grading evaluation and construction method.

The invention provides a novel heterogeneous mixed traffic flow model grading evaluation and construction method, which comprises the following steps:

firstly, establishing function grading evaluation indexes of the automatic driving automobiles, and carrying out unified grading evaluation on the automobiles with different automatic driving functions;

secondly, designing a decision control algorithm for the automatic driving function corresponding to each stage according to the grading index, and generating a multi-grade automatic driving traffic model so as to reflect the running conditions of vehicles with different automatic driving functions;

thirdly, constructing a mixed traffic flow model, designing a driving model distribution algorithm based on micro-element grid weight distribution, and dispersing and uniformly distributing driving models in the second step with different proportions into roads to construct a heterogeneous mixed traffic flow model.

The specific method of the first step is as follows:

based on the interaction between the automatic driving vehicle and the surrounding vehicles and the decision control capability thereof, three grading bases are provided as a perception range, a control function and an understanding depth, and the three grading bases are as follows:

the perception range is road area and traffic vehicle information which can be perceived by the vehicle; the control function is a control capability that the vehicle can make; the understanding depth is the cognition of the vehicle to the peripheral behavior and the road environment and the active response capability of the vehicle to the peripheral behavior;

according to the three grading evaluation bases, the automatic driving vehicles are classified into 4 grades, the grades are from low to high, the interactivity between the vehicles and the peripheral vehicles is sequentially enhanced, and meanwhile, the automatic driving vehicle can adapt to complex traffic games, reasonably predictable real actions can be made, and the traffic passing efficiency is improved;

the heterogeneous traffic model classification is as follows:

the 0-level traffic vehicle model can only sense the information of vehicles in front of the same lane, can only carry out longitudinal driving control, does not understand the traffic environment, and can only passively react;

the level 1 traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, and compared with the level 0 traffic vehicle model, the longitudinal decision is more accurate and quicker, the active judgment is carried out on whether the side lane vehicles cut into the interaction, only the longitudinal driving control can be carried out, and the preliminary understanding is carried out on the traffic environment;

the 2-level traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, and compared with the 1-level traffic vehicle model, the longitudinal decision is more accurate and quicker, the active judgment is carried out on whether the side lane vehicles cut into the interaction, the active response can be carried out on the side lane vehicles, the transverse and longitudinal control can be carried out, and the preliminary understanding is carried out on the traffic environment;

the 3-level traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, compared with the 2-level traffic vehicle, can realize horizontal and longitudinal integrated control, enhances the continuity and the inherent association of driving behaviors, can adaptively learn the optimal decision behaviors, has strong interaction with surrounding vehicles, and has high-level understanding on traffic environment.

The specific method of the second step is as follows:

step 1, aiming at a 0-level traffic vehicle, considering that a longitudinal model needs to deal with the problem of following a front vehicle, dynamically adjusting the acceleration of the vehicle, and keeping a safe and comfortable driving distance, and taking an intelligent driver model (IDM model) as the 0-level traffic vehicle model, wherein the structure is shown in a formula (1):

the expected following distance of the main vehicle for avoiding collision is:

wherein a is _max Maximum acceleration of the main vehicle, delta is an acceleration index, v _t For the current speed of the host vehicle v _e For the expected speed of the host vehicle, deltav _t Is the relative speed of the main car and the front car, s _t B is comfortable deceleration, T is reaction time and s is the relative distance between the main vehicle and the front vehicle _e Is a safe parking space;

step 2, aiming at a level 1 traffic model, considering that a longitudinal decision model needs to be capable of dynamically adjusting the self-vehicle acceleration according to the speed position information of a front vehicle, and keeping a safe and comfortable driving distance; meanwhile, the method needs to adapt to more complex traffic scenes, realizes accurate and flexible decision strategies, designs a longitudinal decision model based on an IDM-LSTM combined vehicle, dynamically adjusts linear combination weights of the IDM and LSTM algorithms through self-adaptive Kalman filtering, and gives consideration to safety and comfort of the IDM algorithm and true accuracy of the LSTM algorithm;

the method comprises the steps that data of vehicles in front and behind running of a main vehicle are collected through vision sensors of a laser radar and a millimeter wave radar arranged on the main vehicle, wherein the data comprise information of speed of the front vehicle, acceleration of the front vehicle, relative speed of the front vehicle and the main vehicle and relative distance between the front vehicle and the main vehicle;

calibrating an IDM model according to the acquired data by using a simulated annealing optimization parameter method, namely an acceleration index delta, a reaction time T and a safe parking interval s _e Comfort deceleration b, desired vehicle speed v _e Maximum speed a _max ；

Constructing an LSTM vehicle following control offline prediction network, firstly extracting the vehicle speed v of the front vehicle from the collected driving data set _f Speed v of main vehicle _h Relative velocity Deltav, acceleration of the front vehicle a _f Acceleration a of host vehicle _h And the relative distance Deltax as an input feature, the desired acceleration a of the host vehicle _exp And a desired following distance v _exp As an output tag;

dividing a data set into a training set, a verification set and a test set, dividing the data into a plurality of sequence samples according to a set time window, constructing an LSTM network model, wherein the LSTM network model comprises an input layer, an LSTM layer, a full-connection layer and an output layer, the input layer receives the 6 features, the LSTM layer is used for extracting time sequence information, the full-connection layer is used for reducing and nonlinear transformation, the multidimensional data output is mapped to two-dimensional data, and the output layer outputs two predicted values;

the method comprises the steps of selecting a mean square error and Adam as a loss function and an optimizer, training and verifying an LSTM network model, adjusting super parameters and a network structure until satisfactory performance is achieved, evaluating the prediction effect of the LSTM network model by using a test set, calculating the root mean square error between a predicted value and a true value, storing the trained LSTM network model as an offline prediction network model as a part of a longitudinal control model of a vehicle, and obtaining the expected following distance and the expected acceleration of a host vehicle through the offline prediction network model by utilizing real-time and historical data acquired by the vehicle in practical application;

in addition, a vector conversion method is designed to judge whether the side lane vehicle cuts into the own lane or not according to the running states of the left and right side lane vehicles and through the running speed of the main vehicle and the running speed and the direction of the side vehicle, so that corresponding action reaction is made in advance;

if the front car is in lane change and cut, the relative distance between the main car and the front car should be calculated again, and the main car will respond to the front car cut in advance, firstly, it needs to judge whether the front car is in lane change or cut, by comparing the longitudinal distance x _p And a lateral distance y _p Is determined by the rate of change of (c) in the database,and->The method adopts a direct distance derivation mode to obtain, considers that the front vehicle possibly changes the track in the direction deviating from the main vehicle, if y is detected _p For a period of time, if the number of the active components increases, the number of the active components is set to->0, when->0, consider that the vehicle is traveling straight on different lanes or changing lane in the direction deviating from the main vehicle, and do not consider the vehicle first, whenIf the distance is not 0, the vehicle is considered to run close to the lane change of the main vehicle, the relative distance is needed to be calculated, and the relative distance x _c The formula is:

wherein the method comprises the steps ofAnd->The calculation formulas are respectively as follows:

the longitudinal decision model makes a corresponding decision action for the behavior of cutting in the front vehicle according to the recalculated relative distance, the behavior reflects that the vehicle has preliminary understanding on the traffic situation of the side lane, and the longitudinal decision model can respond according to the behavior of the side vehicle;

step 3, aiming at a 2-level traffic vehicle model, the longitudinal control decision is the same as that of a 1-level traffic vehicle model, a finite state machine based on rules is used as a transverse lane change decision model, and the process of interaction between a vehicle and surrounding vehicles is simulated through the preset lane change rules, wherein the process is as follows:

when a front vehicle exists in front of the main vehicle, the main vehicle enters a following state, and is switched to a lane changing state when the following state is kept for a set time and the expected vehicle speed is not reached, lane changing preparation is carried out if the surrounding lanes meet better driving conditions, and lane changing conditions are judged;

taking safety factors into consideration in the lane changing condition, selecting a TTC index based on a safety distance model, if a target lane has a front and a rear vehicle, if the speed of a main vehicle is lower than the speed of a rear vehicle, or if the speed of the main vehicle is higher than the speed of the front vehicle, performing TTC judgment, and if the TTC is higher than a threshold, performing lane changing operation;

in the course of changing lane, if the acceleration of rear car exceeds 1m/s in the course of changing lane ² Based on safety consideration, waiting for the channel change opportunity again, so that the driving safety is improved;

and 4, aiming at a 3-level traffic model, considering that the vehicle has higher intelligence and can have deep understanding on traffic situation, taking a deep reinforcement learning algorithm DQN as a decision control model, performing longitudinal control and integrated control of lane change decision, and enhancing continuity and internal association of driving behaviors and interaction with surrounding vehicles compared with horizontal and longitudinal decoupling control of a 2-level model. Defining a state space, an action space and a reward function, wherein the state space comprises information of positions, speeds, accelerations and course angles of a vehicle and surrounding vehicles, the action space comprises two dimensions of longitudinal acceleration and transverse lane change, and the reward function comprehensively considers factors of running efficiency, safety and comfort;

constructing a DQN network model, wherein the DQN network model comprises a Q network and a target network, the Q network is used for estimating the Q value of each action in the current state, the target network is used for estimating the maximum Q value of the next state, the structures of the two networks can be composed of a plurality of full-connection layers, the dimension of an input layer is the size of a state space, and the dimension of an output layer is the size of an action space;

selecting a mean square error and Adam as a loss function and an optimizer, training a DQN network model, exploring and utilizing an environment by using an epsilon-greedy strategy, storing and sampling transfer data by using an experience playback mechanism, and updating parameters of a target network by using a fixed frequency in the training process;

evaluating the control effect of the DQN network model by using the test set, and calculating error indexes, such as root mean square error and average absolute error, between the predicted action and the real action;

the trained DQN network model is stored as an integrated control network model, can be used in the scenes of vehicle following and lane changing, and in practical application, the optimal longitudinal acceleration and transverse lane change are obtained through the integrated control network model by utilizing real-time and historical data acquired by a vehicle-to-vehicle communication technology.

The specific method of the third step is as follows:

step 1, dispersing a road space into micro-element grids with different sizes, wherein the size of a single grid is Deltax multiplied by Deltay, and Deltax and Deltay are jointly determined according to the width of a road lane and the size of a vehicle;

dividing the road space into a plurality of strip-shaped areas with length deltax along the road direction, and dividing each strip-shaped area into a plurality of small areas with width deltay along the direction perpendicular to the road, namely grids, wherein the area of each grid is deltax deltay, specifically deltax is greater than or equal to the length of the vehicle, and deltay is greater than or equal to the width of the vehicle;

the mesh division aims at dispersing the road space into a plurality of areas, wherein each area represents the position and the state of a traffic vehicle; the principle of grid division is that the size of each area is matched with the size of a vehicle, and the subsequent optimizing efficiency is influenced by avoiding excessive grid number;

step 2, defining an objective function f to represent the distribution conditions of different grades of traffic vehicles:

where n is the number of traffic models, w _ij Is the weight between the ith and jth traffic models, d _ij Distance between the ith and jth traffic models;

w _ij depending on the type of driver:

if the ith and jth traffic class are the same, i.e. l _i ＝l _j Then w _ij -1, avoiding the same level traffic model to be distributed to adjacent areas;

if the ith and jth drivers are of different types, i.e./ _i ≠l _j Then w _ij The specific value of =1/1.2/1.4 depends on the difference of the traffic classes, so as to increase the mixing degree of the traffic models of different classes.

d _ij The calculation formula of (2) is as follows:

step 3, searching an optimal traffic vehicle model distribution scheme by using a particle swarm optimization algorithm, initializing a group of particles with random number N, wherein each particle represents a possible traffic vehicle model distribution scheme, namely the coordinate position of a grid where each traffic vehicle is located, and algorithm parameters in a two-dimensional search space are respectively as follows:

X _id ＝(x _ix ,x _iy ) (9)

V _id ＝(v _ix ,v _iy ) (10)

P _id,best ＝(p _ix ,p _iy ) (11)

P _d,gbest ＝(p _x,gbest ,p _y,gbest ) (12)

wherein X is _id For the position of the ith particle, V _id Is the speed of the ith particle and includes the distance and direction of particle movement, P _id，best For the optimal position of the ith particle, P _d，best Searching the optimal position for the group;

step 4, updating the particle speed and the position:

wherein k is the iteration number, w is the inertial weight, c ₁ For individual learning factors, c ₂ R is a group learning factor ₁ ，r ₂ Is interval [0,1 ]]Increasing the randomness of the search;

and 5, repeatedly implementing the steps, calculating the objective function value f (x) of each particle, updating the speed and the position of each particle until the termination condition is met, namely, the iteration times are greater than the set maximum iteration times by 50 times, outputting the global optimal solution g, namely, the optimal traffic model distribution scheme, and further completing the construction of the heterogeneous mixed traffic flow model.

The invention has the beneficial effects that:

the novel grading evaluation and construction method for the heterogeneous mixed traffic flow model provided by the invention carries out unified grading evaluation on the automatic driving vehicles in the heterogeneous mixed traffic flow by establishing the multidimensional grading evaluation index, establishes the grading index which better accords with traffic flow characteristics, and simultaneously realizes uniform distribution and dispersion of the automatic driving vehicles with different functions, namely different grades, in a given road, so as to construct the heterogeneous mixed traffic flow simulation model. The method can be applied to heterogeneous mixed traffic flow models with high confidence coefficient for intelligent automobile simulation research, can be popularized and applied to research in the fields of intelligent automobile decision planning, control, testing and the like, and has wide application prospects.

Drawings

Fig. 1 is a schematic diagram of a hierarchical evaluation and construction flow of a heterogeneous mixed traffic flow model according to the present invention.

Fig. 2 is a flow chart of an IDM-LSTM network based on adaptive kalman filtering according to the present invention.

Fig. 3 is a schematic diagram of a side car cutting in the present invention.

Fig. 4 is a schematic diagram of a state machine implementing a lateral channel switching strategy according to the present invention.

Fig. 5 is a flow chart of distribution of traffic models based on infinitesimal mesh weight distribution according to the invention.

Detailed Description

Please refer to fig. 1 to 5:

the invention provides a novel heterogeneous mixed traffic flow model grading evaluation and construction method, which comprises the following steps:

the method comprises the following steps of firstly, establishing function grading evaluation indexes of the automatic driving vehicles, and carrying out unified grading evaluation on the automobiles with different automatic driving functions, wherein the method comprises the following specific steps:

the different functions of the automatic driving vehicle can be represented as the adaptation and cooperation degree and range of the vehicle in a complex traffic environment in traffic flow, and can be reflected on the interaction game with the vehicle and Zhou Che and the autonomous decision making capability of the vehicle, so that three grading bases are provided on the basis of the interaction of the automatic driving vehicle and surrounding vehicles and the decision making control capability thereof, namely a perception range, a control function and an understanding depth.

Specifically, the perception range is road area and traffic vehicle information that can be perceived by the vehicle; the control function is a control capability that the vehicle can make; the depth of understanding is the awareness of the vehicle to the surrounding behavior and road environment and the active response capability of the vehicle to the surrounding behavior.

According to the three above-mentioned grading evaluation bases, the automatically driven vehicles were classified into 4 grades, as shown in table 1. The grade is from low to high, the interactivity of vehicles and peripheral vehicles is enhanced in sequence, and meanwhile, the method can adapt to complex traffic games, reasonably foreseeable real actions can be made, and traffic passing efficiency is improved.

TABLE 1 heterogeneous traffic model classification

Specifically, the 0-level traffic model can only sense the information of vehicles in front of the same lane, can only perform longitudinal driving control, is not understood to the traffic environment, and can only passively react.

The level 1 traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, and compared with the level 0 traffic vehicle model, the longitudinal decision is more accurate and rapid, the active judgment is carried out on whether the side lane vehicles cut into the interaction, the longitudinal driving control can only be carried out, and the preliminary understanding is provided for the traffic environment.

The 2-level traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, and compared with the 1-level traffic vehicle model, the longitudinal decision is more accurate and rapid, the active judgment is carried out on whether the side lane vehicles cut into the interaction, the active response is carried out on the side lane vehicles, the transverse and longitudinal control can be carried out, and the preliminary understanding is carried out on the traffic environment.

The 3-level traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, can realize horizontal and longitudinal integrated control compared with the 2-level traffic vehicle, enhances the continuity and internal association of driving behaviors, can adaptively learn the optimal decision behaviors, has strong interaction with surrounding vehicles, and has high-level understanding on traffic environment.

Secondly, designing a decision control algorithm for the automatic driving function corresponding to each stage according to the grading index, and generating a multi-grade automatic driving traffic model so as to reflect the running conditions of vehicles with different automatic driving functions;

aiming at the 0-level traffic vehicle, considering the problem that the longitudinal model needs to process the following of the front vehicle, the self-vehicle acceleration is dynamically adjusted, and the safe and comfortable driving distance is kept, because the intelligent driver model (IDM model) can better simulate the following behaviors of drivers and the front vehicle in different traffic flows, and has good performance, stable and smooth following can be realized, the IDM model is used as the 0-level traffic vehicle model. The structure is shown in formula (1):

further, the formula can be divided into two parts,expressing acceleration of the vehicle traveling on a non-congested road,/->Indicating the interactive braking acceleration of the host vehicle as it approaches the preceding vehicle.

The expected following distance of the main vehicle for avoiding collision is:

wherein a is _max Maximum acceleration of the main vehicle, delta is an acceleration index, v _t For the current speed of the host vehicle v _e For the expected speed of the host vehicle, deltav _t Is the relative speed of the main car and the front car, s _t B is comfortable deceleration, T is reaction time and s is the relative distance between the main vehicle and the front vehicle _e Is a safe parking space.

Aiming at a level 1 traffic model, considering that a longitudinal decision model needs to be capable of dynamically adjusting the self-vehicle acceleration according to the speed position information of a front vehicle, and keeping safe and comfortable driving distance; meanwhile, the method needs to adapt to more complex traffic scenes and realizes accurate and flexible decision strategies. The longitudinal decision model of the combined vehicle based on the IDM-LSTM is designed, referring to fig. 2, the linear combination weights of the IDM and LSTM are dynamically adjusted through the adaptive Kalman filtering, the safety and the comfort of the IDM algorithm are considered, and the real accuracy of the LSTM algorithm is considered. Specifically, the state vector is the acceleration difference value output by the IDM and LSTM algorithm, the observation vector is the output acceleration of the IDM and LSTM, the acceleration difference value is optimally estimated through the self-adaptive Kalman filtering, and then the obtained result is added with the predicted acceleration of the LSTM to obtain the final output acceleration, namely:

wherein the method comprises the steps ofIs a posterior estimate of the acceleration difference obtained by adaptive Kalman filtering.

The method comprises the steps that data of vehicles in front and behind running of a main vehicle are collected through vision sensors such as a laser radar, a millimeter wave radar and the like arranged on the main vehicle, wherein the data comprise information such as speed of the front vehicle, acceleration of the front vehicle, relative speed of the front vehicle and the main vehicle, relative distance between the front vehicle and the main vehicle and the like;

calibrating an IDM model according to the acquired data by using a simulated annealing optimization parameter method, namely an acceleration index delta, a reaction time T and a safe parking interval s _e Comfort deceleration b, desired vehicle speed v _e Maximum speed a _max . Firstly, randomly generating an initial value for parameters to be calibrated, and randomly generating the initial value for the parameters at presentAnd continuously randomly generating a new parameter solution in the field of the solution, calculating an objective function value of the parameter solution, selecting the new solution as the current solution if the value of the objective function of the new solution is smaller than the objective function value of the current parameter solution, and repeatedly iterating the process until the iteration times are reached and the termination condition is met, and then obtaining the optimal solution.

Constructing an LSTM vehicle following control offline prediction network, firstly extracting the vehicle speed v of the front vehicle from the collected driving data set _f Speed v of main vehicle _h Relative velocity Deltav, acceleration of the front vehicle a _f Acceleration a of host vehicle _h And the relative distance Deltax as an input feature, the desired acceleration a of the host vehicle _exp And a desired following distance v _exp As an output tag.

And dividing the data set into a training set, a verification set and a test set, and dividing the data into a plurality of sequence samples according to a certain time window. And constructing an LSTM network model, wherein the LSTM network model comprises an input layer, an LSTM layer, a full connection layer and an output layer. The input layer receives the 6 features, the LSTM layer is used for extracting time sequence information, the full connection layer is used for reducing and nonlinear transformation, the multidimensional data output is mapped to two-dimensional data, and the output layer outputs two predicted values.

Specifically, [ v ] _f ,v _h ,Δv,a _f ,a _h ,Δx]Normalization as input x of model _t Through LSTM layer, full connection layer and output layer, through inverse normalization output [ a ] _exp ,x _exp ]。

And selecting the mean square error and Adam as a loss function and an optimizer, training and verifying the LSTM network model, and adjusting the super-parameters and the network structure until the satisfactory performance is achieved. And evaluating the prediction effect of the LSTM network model by using the test set, and calculating the root mean square error between the predicted value and the true value. The trained LSTM network model is saved as an offline predictive network model as part of the vehicle longitudinal control model. In practical application, the expected following distance and the expected acceleration of the host vehicle are obtained through an offline prediction network model by utilizing real-time and historical data acquired by the vehicle.

In addition, the vector conversion method is designed to judge whether the side lane vehicle cuts into the own lane or not according to the running states of the left and right side lane vehicles by the running speed of the own vehicle and the running speed and direction of the side vehicle, so that corresponding action reaction is made in advance, and refer to fig. 3.

Specifically, when the front car is in the lane change cut, the relative distance between the main car and the front car should be recalculated, and the main car can respond to the front car cut in advance. Firstly, judging whether the front car is changed or cut in, and comparing longitudinal distance x _p And a lateral distance y _p Is determined by the rate of change of (c) in the database,and->The method adopts a direct distance derivation mode to obtain, considers that the front vehicle possibly changes the track in the direction deviating from the main vehicle, if y is detected _p For a period of time, if the number of the active components increases, the number of the active components is set to->Is 0. When->For 0, it is considered that the vehicle is traveling straight on a different lane or is changing lane in a direction deviating from the main vehicle, and is not considered first. When->If the distance is not 0, the vehicle can be considered to run close to the lane change of the main vehicle, and the relative distance x needs to be calculated _c The formula is:

wherein the method comprises the steps ofAnd->The calculation formulas are respectively as follows:

the longitudinal decision model makes a corresponding decision action for the behavior of cutting into the front vehicle according to the recalculated relative distance, the behavior reflects that the main vehicle has preliminary understanding on the traffic situation of the side lane, and the longitudinal decision model can respond according to the behavior of the side vehicle.

For the 2-level traffic model, the longitudinal control decision is the same as that of the 1-level traffic model, a rule-based finite state machine is used as a transverse lane change decision model, and the process of vehicle interaction with surrounding vehicles is simulated through the preset lane change rule, and the method is shown in fig. 4. The process comprises the following steps:

when a front vehicle exists in front of the main vehicle, the main vehicle enters a following state, and if the front vehicle drives away during the following process, the main vehicle keeps a cruising state. When the following state is kept for a certain time and the expected speed is not reached, the lane change intention is generated to switch to the lane change intention state, lane change preparation is carried out if the surrounding lanes meet better driving conditions, lane change condition is judged, the lanes are replaced if the lane change condition is met, and the state of the vehicle before the following is kept continuously if the lane change condition is not met.

The lane change condition considers safety factors, TTC indexes based on a safety distance model are selected, if a target lane has front and rear vehicles, TTC judgment is carried out when the speed of the vehicle is lower than the speed of the rear vehicle or the speed of the vehicle is higher than the speed of the front vehicle, and if the TTC is higher than a threshold value, the lane change operation is carried out.

In the course of changing lane, if the acceleration of rear car exceeds 1m/s in the course of changing lane ² Then based onAnd in the safety consideration, the lane change opportunity is waited again, and the driving safety is improved.

Aiming at a 3-level traffic model, the vehicle has higher intelligence, can be deeply understood on traffic situation, takes a deep reinforcement learning algorithm DQN as a decision control model, performs integrated control of longitudinal control and lane change decision, and has the advantages of enhanced continuity and internal association of driving behavior and enhanced interaction with surrounding vehicles compared with 2-level model transverse and longitudinal decoupling control. A state space, an action space, and a reward function are defined. The state space comprises information such as the positions, the speeds, the accelerations, the course angles and the like of the self-vehicle and surrounding vehicles, the action space comprises two dimensions of longitudinal acceleration and transverse lane change, and the reward function comprehensively considers factors such as running efficiency, safety, comfort and the like.

The DQN network model is constructed to include a Q network and a target network. The Q network is used for estimating the Q value of each action in the current state, and the target network is used for estimating the maximum Q value of the next state. The structure of the two networks can be composed of a plurality of full-connection layers, the dimension of the input layer is the size of the state space, and the dimension of the output layer is the size of the action space.

And selecting the mean square error and Adam as a loss function and an optimizer to train the DQN network model. In the training process, an epsilon-greedy strategy is used for exploring and utilizing the environment, an empirical playback mechanism is used for storing and sampling transfer data, and a fixed frequency is used for updating parameters of a target network.

The control effect of the DQN network model is evaluated using the test set, and error indicators, such as root mean square error and mean absolute error, between the predicted action and the actual action are calculated.

The trained DQN network model is stored as an integrated control network model and can be used in the scenes of vehicle following and lane changing. In practical application, the optimal longitudinal acceleration and transverse lane change are obtained through an integrated control network model by utilizing real-time and historical data acquired by a vehicle-to-vehicle communication technology.

Thirdly, constructing a mixed traffic flow model, designing a driving model distribution algorithm based on micro-element grid weight distribution, and dispersing and uniformly distributing driving models in different quantity and set proportion into roads to construct a heterogeneous mixed traffic flow model, wherein the figure 5 is referred to.

First, the road space is discretized into microcell grids of different sizes, and the size of a single grid is deltax deltay, wherein deltax and deltay are jointly determined according to the road lane width and the size of the vehicle.

The road space is firstly divided into a plurality of strip-shaped areas with the length delta x along the road direction, and then each strip-shaped area is divided into a plurality of small areas with the width delta y along the direction perpendicular to the road, namely grids. Thus, the area of each grid is Δx×Δy. Specifically, Δx should be greater than or equal to the length of the vehicle and Δy should be greater than or equal to the width of the vehicle.

The mesh division aims at dispersing the road space into a plurality of areas, wherein each area represents the position and the state of a traffic vehicle; the principle of grid division is to enable the size of each area to be matched with the size of a vehicle, and the influence on the subsequent optimizing efficiency can be avoided by avoiding excessive grid number.

Then, an objective function f is defined to represent the distribution situation of different levels of traffic vehicles:

where n is the number of traffic models, w _ij Is the weight between the ith and jth traffic models, d _ij Is the distance between the ith and jth traffic models.

w _ij Depending on the type of driver:

if the ith and jth traffic class are the same, i.e. l _i ＝l _j Then w _ij = -1, avoiding the same level traffic model to be distributed to neighboring areas.

If the ith and jth drivers are of different types, i.e./ _i ≠l _j Then w _ij The specific value of =1/1.2/1.4 depends on the difference of the traffic classes, so as to increase the mixing degree of the traffic models of different classes.

d _ij The calculation formula of (2) is as follows:

searching an optimal traffic vehicle model distribution scheme by using a particle swarm optimization algorithm, initializing a group of particles with random number N, wherein each particle represents a possible traffic vehicle model distribution scheme, namely the coordinate position of a grid where each traffic vehicle is located, and algorithm parameters in a two-dimensional search space are respectively as follows:

X _id ＝(x _ix ,x _iy ) (23)

V _id ＝(v _ix ,v _iy ) (24)

P _id,best ＝(p _ix ,p _iy ) (25)

P _d,gbest ＝(p _x,gbest ,p _y,gbest ) (26)

wherein X is _id For the position of the ith particle, V _id For the speed of the ith particle (including distance and direction of particle movement), P _id，best For the optimal position of the ith particle, P _d，best The best location searched for the population.

The particle velocity and position are then updated:

specifically, the speed update may be composed of an inertia part, a cognitive part, and a social part, and may be interpreted as a moving direction=an inertia direction+an individual optimal direction+a population optimal direction of the next iteration of the particle. Wherein k is the iteration number, w is the inertial weight, c ₁ For individual learning factors, c ₂ R is a group learning factor ₁ ，r ₂ Is interval [0,1 ]]To increase the randomness of the search.

And finally, repeatedly implementing the steps, calculating the objective function value f (x) of each particle, and updating the speed and the position of each particle until the termination condition is met, namely the iteration times are greater than the set maximum iteration times. And outputting a global optimal solution g, namely an optimal traffic model distribution scheme, and further completing the construction of a heterogeneous mixed traffic flow model.

In addition, the specific features described in the above embodiments may be combined in any suitable manner, and in order to avoid unnecessary repetition, various possible combinations are not described further.

Claims (4)

1. A novel heterogeneous mixed traffic flow model grading evaluation and construction method is characterized in that: the method comprises the following steps:

firstly, establishing function grading evaluation indexes of the automatic driving vehicles, and carrying out unified grading evaluation on the vehicles with different automatic driving functions;

secondly, designing a decision control algorithm for the automatic driving function corresponding to each stage according to the grading index, and generating a multi-grade automatic driving traffic model so as to reflect the running conditions of vehicles with different automatic driving functions;

thirdly, constructing a mixed traffic flow model, designing a driving model distribution algorithm based on micro-element grid weight distribution, and dispersing and uniformly distributing driving models in the second step with different proportions into roads to construct a heterogeneous mixed traffic flow model.

2. The method for grading evaluation and construction of a novel heterogeneous mixed traffic flow model according to claim 1, which is characterized in that: the specific method of the first step is as follows:

based on the interaction between the automatic driving vehicle and the surrounding vehicles and the decision control capability thereof, three grading bases are provided as a perception range, a control function and an understanding depth, and the three grading bases are as follows:

the perception range is road area and traffic vehicle information which can be perceived by the vehicle; the control function is a control capability that the vehicle can make; the understanding depth is the cognition of the vehicle to the peripheral behavior and the road environment and the active response capability of the vehicle to the peripheral behavior;

according to the three grading evaluation bases, the automatic driving vehicles are classified into 4 grades, the grades are from low to high, the interactivity between the vehicles and the peripheral vehicles is sequentially enhanced, and meanwhile, the automatic driving vehicle can adapt to complex traffic games, reasonably predictable real actions can be made, and the traffic passing efficiency is improved;

the heterogeneous traffic model classification is as follows:

the 0-level traffic vehicle model can only sense the information of vehicles in front of the same lane, can only carry out longitudinal driving control, does not understand the traffic environment, and can only passively react;

the level 1 traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, and compared with the level 0 traffic vehicle model, the longitudinal decision is more accurate and quicker, the active judgment is carried out on whether the side lane vehicles cut into the interaction, only the longitudinal driving control can be carried out, and the preliminary understanding is carried out on the traffic environment;

the 2-level traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, and compared with the 1-level traffic vehicle model, the longitudinal decision is more accurate and quicker, the active judgment is carried out on whether the side lane vehicles cut into the interaction, the active response can be carried out on the side lane vehicles, the transverse and longitudinal control can be carried out, and the preliminary understanding is carried out on the traffic environment;

the 3-level traffic vehicle model can sense the vehicle information of the same lane and left and right side lanes, compared with the 2-level traffic vehicle, can realize horizontal and longitudinal integrated control, enhances the continuity and the inherent association of driving behaviors, can adaptively learn the optimal decision behaviors, has strong interaction with surrounding vehicles, and has high-level understanding on traffic environment.

3. The method for grading evaluation and construction of a novel heterogeneous mixed traffic flow model according to claim 1, which is characterized in that: the specific method of the second step is as follows:

step 1, aiming at a 0-level traffic vehicle, considering that a longitudinal model needs to deal with the problem of following a front vehicle, dynamically adjusting the acceleration of a vehicle of the vehicle, and keeping a safe and comfortable driving distance, taking an intelligent driver model (IDM model) as the 0-level traffic vehicle model, wherein the structure is shown in a formula (1):

the expected following distance of the main vehicle for avoiding collision is:

wherein a is _max Maximum acceleration of the main vehicle, delta is an acceleration index, v _t For the current speed of the host vehicle v _e The expected speed of the main vehicle is Deltavt, the relative speed of the main vehicle and the front vehicle is s _t B is comfortable deceleration, T is reaction time and s is the relative distance between the main vehicle and the front vehicle _e Is a safe parking space;

step 2, aiming at a level 1 traffic model, considering that a longitudinal decision model needs to be capable of dynamically adjusting the self-vehicle acceleration according to the speed position information of a front vehicle, and keeping a safe and comfortable driving distance; meanwhile, the method needs to adapt to more complex traffic scenes, realizes accurate and flexible decision strategies, designs a longitudinal decision model based on an IDM-LSTM combined vehicle, dynamically adjusts linear combination weights of the IDM and LSTM algorithms through self-adaptive Kalman filtering, and gives consideration to safety and comfort of the IDM algorithm and true accuracy of the LSTM algorithm;

the method comprises the steps that data of vehicles in front and behind running of a main vehicle are collected through vision sensors of a laser radar and a millimeter wave radar arranged on the main vehicle, wherein the data comprise information of speed of the front vehicle, acceleration of the front vehicle, relative speed of the front vehicle and the main vehicle and relative distance between the front vehicle and the main vehicle;

calibrating an IDM model according to the acquired data by using a simulated annealing optimization parameter method, namely an acceleration index delta, a reaction time T and a safe parking interval s _e Comfort deceleration b, desired vehicle speed v _e Maximum speed a _max ；

Constructing an LSTM vehicle following control offline prediction network, firstly extracting the vehicle speed v of the front vehicle from the collected driving data set _f Speed v of main vehicle _h Relative velocity Deltav, acceleration of the front vehicle a _f Acceleration a of host vehicle _h And the relative distance Deltax as an input feature, the desired acceleration a of the host vehicle _exp And a desired following distance v _exp As an output tag;

dividing a data set into a training set, a verification set and a test set, dividing the data into a plurality of sequence samples according to a set time window, constructing an LSTM network model, wherein the LSTM network model comprises an input layer, an LSTM layer, a full-connection layer and an output layer, the input layer receives the 6 features, the LSTM layer is used for extracting time sequence information, the full-connection layer is used for reducing and nonlinear transformation, the multidimensional data output is mapped to two-dimensional data, and the output layer outputs two predicted values;

the method comprises the steps of selecting a mean square error and Adam as a loss function and an optimizer, training and verifying an LSTM network model, adjusting super parameters and a network structure until satisfactory performance is achieved, evaluating the prediction effect of the LSTM network model by using a test set, calculating the root mean square error between a predicted value and a true value, storing the trained LSTM network model as an offline prediction network model as a part of a longitudinal control model of a vehicle, and obtaining the expected following distance and the expected acceleration of a host vehicle through the offline prediction network model by utilizing real-time and historical data acquired by the vehicle in practical application;

in addition, a vector conversion method is designed to judge whether the side lane vehicle cuts into the own lane or not according to the running states of the left and right side lane vehicles and through the running speed of the own vehicle and the running speed and the direction of the side vehicle, so that corresponding action reaction is made in advance;

if the front car is in lane change and cut, the relative distance between the main car and the front car should be calculated again, and the main car will respond to the front car cut in advance, firstly, it needs to judge whether the front car is in lane change or cut, by comparing the longitudinal distance x _p And a lateral distance y _p Is determined by the rate of change of (c) in the database,and->The method adopts a direct distance derivation mode to obtain, considers that the front vehicle possibly changes the track in the direction deviating from the main vehicle, if y is detected _p For a period of time, if the number of the active components increases, the number of the active components is set to->Is 0 whenIf 0, the vehicle is considered to run straight on different lanes or change lane in the direction deviating from the main vehicle, and is not considered first, when +.>If the distance is not 0, the vehicle is considered to run close to the lane change of the main vehicle, the relative distance is needed to be calculated, and the relative distance x _c The formula is:

wherein the method comprises the steps ofAnd->The calculation formulas are respectively as follows:

the longitudinal decision model makes a corresponding decision action for the behavior of cutting in the front vehicle according to the recalculated relative distance, the behavior reflects that the vehicle has preliminary understanding on the traffic situation of the side lane, and the longitudinal decision model can respond according to the behavior of the side vehicle;

step 3, aiming at a 2-level traffic vehicle model, the longitudinal control decision is the same as that of a 1-level traffic vehicle model, a finite state machine based on rules is used as a transverse lane change decision model, and the process of interaction between a vehicle and surrounding vehicles is simulated through the preset lane change rules, wherein the process is as follows:

when a front vehicle exists in front of the vehicle, the main vehicle enters a following state, and is switched to a lane changing state when the following state is kept for a set time and the expected vehicle speed is not reached, lane changing preparation is carried out if the surrounding lanes meet better driving conditions, and lane changing conditions are judged;

taking safety factors into consideration in the lane changing condition, selecting a TTC index based on a safety distance model, if a target lane has a front vehicle and a rear vehicle, if the speed of the host vehicle is lower than the speed of the rear vehicle or the speed of the host vehicle is higher than the speed of the front vehicle, performing TTC judgment, and if the TTC is higher than a threshold value, performing lane changing operation;

in the course of changing lane, if the acceleration of rear car exceeds 1m/s in the course of changing lane ² Based on safety consideration, waiting for the channel change opportunity again, so that the driving safety is improved;

step 4, aiming at a 3-level traffic model, considering that the vehicle has higher intelligence and can have deep understanding on traffic situation, taking a deep reinforcement learning algorithm DQN as a decision control model, performing integrated control of longitudinal control and lane change decision, compared with horizontal and longitudinal decoupling control of a 2-level model, enhancing continuity and internal association of driving behavior, enhancing interaction with surrounding vehicles, defining a state space, an action space and a reward function, wherein the state space comprises information of positions, speeds, accelerations and heading angles of the vehicle and surrounding vehicles, the action space comprises two dimensions of longitudinal acceleration and transverse lane change, and the reward function comprehensively considers factors of driving efficiency, safety and comfort;

constructing a DQN network model, wherein the DQN network model comprises a Q network and a target network, the Q network is used for estimating the Q value of each action in the current state, the target network is used for estimating the maximum Q value of the next state, the structures of the two networks can be composed of a plurality of full-connection layers, the dimension of an input layer is the size of a state space, and the dimension of an output layer is the size of an action space;

selecting a mean square error and Adam as a loss function and an optimizer, training a DQN network model, exploring and utilizing an environment by using an epsilon-greedy strategy, storing and sampling transfer data by using an experience playback mechanism, and updating parameters of a target network by using a fixed frequency in the training process;

evaluating the control effect of the DQN network model by using the test set, and calculating error indexes, such as root mean square error and average absolute error, between the predicted action and the real action;

the trained DQN network model is stored as an integrated control network model, can be used in the scenes of vehicle following and lane changing, and in practical application, the optimal longitudinal acceleration and transverse lane change are obtained through the integrated control network model by utilizing real-time and historical data acquired by a vehicle-to-vehicle communication technology.

4. The method for grading evaluation and construction of a novel heterogeneous mixed traffic flow model according to claim 1, which is characterized in that: the specific method of the third step is as follows:

step 1, dispersing a road space into micro-element grids with different sizes, wherein the size of a single grid is Deltax multiplied by Deltay, and Deltax and Deltay are jointly determined according to the width of a road lane and the size of a vehicle;

dividing the road space into a plurality of strip-shaped areas with length deltax along the road direction, and dividing each strip-shaped area into a plurality of small areas with width deltay along the direction perpendicular to the road, namely grids, wherein the area of each grid is deltax deltay, specifically deltax is greater than or equal to the length of the vehicle, and deltay is greater than or equal to the width of the vehicle;

the mesh division aims at dispersing the road space into a plurality of areas, wherein each area represents the position and the state of a traffic vehicle; the principle of grid division is that the size of each area is matched with the size of a vehicle, and the subsequent optimizing efficiency is influenced by avoiding excessive grid number;

step 2, defining an objective function f to represent the distribution conditions of different grades of traffic vehicles:

where n is the number of traffic models, w _ij Is the weight between the ith and jth traffic models, d _ij Distance between the ith and jth traffic models;

w _ij depending on the type of driver:

if the ith and jth traffic class are the same, i.e. l _i ＝l _j Then w _jj -1, avoiding the same level traffic model to be distributed to adjacent areas;

if the ith and jth drivers are of different types, i.e./ _i ≠l _j Then w _ij The specific value of the combination of the two classes of traffic models is 1/1.2/1.4, and depends on the difference value of the classes of traffic, so that the mixing degree of the models of the traffic models of different classes is increased;

d _ij the calculation formula of (2) is as follows:

step 3, searching an optimal traffic vehicle model distribution scheme by using a particle swarm optimization algorithm, initializing a group of particles with random number N, wherein each particle represents a possible traffic vehicle model distribution scheme, namely the coordinate position of a grid where each traffic vehicle is located, and algorithm parameters in a two-dimensional search space are respectively as follows:

X _id ＝(x _ix ,x _iy ) (9)

V _id ＝(v _ix ,v _iy ) (10)

P _id,best ＝(p _ix ,p _iy ) (11)

P _d,gbest ＝(p _x,gbest ,p _y,gbest ) (12)

wherein X is _id For the position of the ith particle, V _id Is the speed of the ith particle and includes the distance and direction of particle movement, P _id，best For the optimal position of the ith particle, P _d，best Searching the optimal position for the group;

step 4, updating the particle speed and the position:

wherein k is the iteration number, w is the inertial weight, c ₁ For individual learning factors, c ₂ R is a group learning factor ₁ ，r ₂ Is interval [0,1 ]]Increasing the randomness of the search;

and 5, repeatedly implementing the steps, calculating the objective function value f (x) of each particle, updating the speed and the position of each particle until the termination condition is met, namely, the iteration times are greater than the set maximum iteration times by 50 times, outputting the global optimal solution g, namely, the optimal traffic model distribution scheme, and further completing the construction of the heterogeneous mixed traffic flow model.