A new hierarchical evaluation and construction method for heterogeneous mixed traffic flow models

Abstract

The present invention discloses a novel hierarchical evaluation and construction method for heterogeneous mixed traffic flow models, which comprises the following steps: first, establishing hierarchical evaluation indicators for autonomous driving vehicle functions; second, designing decision control algorithms for autonomous driving functions corresponding to each level according to the hierarchical indicators; and third, constructing a mixed traffic flow model. Beneficial effects: A hierarchical indicator that is more in line with traffic flow characteristics is established, while achieving uniform distribution of autonomous driving vehicles of different functions, i.e., different levels, on a given road, thereby constructing a heterogeneous mixed traffic flow simulation model. This method can be applied to provide an effective and highly confident heterogeneous mixed traffic flow model for intelligent vehicle simulation research, and can be extended to research in the fields of intelligent vehicle decision planning, control, testing, etc., and has broad application prospects.

Description

Technical Field

[0001] The present invention relates to a model hierarchical evaluation and construction method, and in particular to a novel heterogeneous mixed traffic flow model hierarchical evaluation and construction method. Background Art

[0002] At present, with the development of intelligent connected vehicles and autonomous driving, the proportion of vehicles with different autonomous driving functions on the road is gradually increasing, making the mutual influence between vehicle behaviors more complex and changeable.

[0003] At present, most mixed traffic flow simulation models are built based on a single model, which makes it difficult to reflect the behavior and impact of vehicles with different autonomous driving functions in traffic flow. There is a lack of a unified framework or evaluation method to evaluate and develop heterogeneous mixed traffic flow models. Heterogeneous mixed traffic flow models refer to traffic flow models that interact and dynamically change between different types of vehicles. For example, CN112329248A discloses a road mixed traffic flow simulation system based on a multi-agent system, but this method does not consider the coordination and conflict between autonomous driving vehicles with different functions, and does not evenly distribute the model.

[0004] It is of great significance to construct a new heterogeneous mixed traffic flow classification evaluation method that incorporates autonomous driving vehicles with different functions and design the corresponding traffic flow model. Summary of the Invention

[0005] The purpose of this invention is to solve the problems that existing traffic flow simulation models cannot truly reflect the traffic situation of mixed autonomous driving vehicles with different functions on the road and are difficult to meet the high-confidence simulation requirements of heterogeneous mixed traffic flows. A new hierarchical evaluation and construction method for heterogeneous mixed traffic flow models is provided.

[0006] The present invention provides a novel hierarchical evaluation and construction method for heterogeneous mixed traffic flow models, which includes the following steps:

[0007] The first step is to establish a functional grading evaluation index for autonomous vehicles, and to uniformly grade and evaluate vehicles with different autonomous driving functions.

[0008] The second step is to design decision-making control algorithms for each level of autonomous driving function according to the grading indicators, generating a multi-level autonomous driving traffic model to reflect the operation of vehicles with different autonomous driving functions.

[0009] The third step is to construct a mixed traffic flow model and design a driving model distribution algorithm based on micro-element grid weight allocation. The driving models described in step 2 of different proportions are evenly distributed on the road to construct a heterogeneous mixed traffic flow model.

[0010] The specific method of the first step is as follows:

[0011] Based on the interaction between autonomous vehicles and surrounding vehicles and their own decision-making and control capabilities, three classification criteria are proposed: perception range, control function, and depth of understanding. The details are as follows:

[0012] The perception range refers to the road area and traffic information that the vehicle can perceive; the control function refers to the control capability of the vehicle; and the understanding depth refers to the vehicle's cognition of the surrounding vehicle behavior and road environment, as well as the vehicle's ability to actively respond to the surrounding vehicle behavior.

[0013] Based on the three aforementioned evaluation criteria, autonomous vehicles are divided into four levels. From low to high, the interactivity of the vehicle with surrounding vehicles increases successively. At the same time, it can adapt to complex traffic games, make reasonable and predictable real actions, and improve traffic efficiency.

[0014] The classification of heterogeneous traffic flow vehicle models is as follows:

[0015]

[0016] The Level 0 traffic vehicle model can only perceive information about vehicles in the same lane ahead and can only perform longitudinal driving control. It has no understanding of the traffic environment and can only react passively.

[0017] The Level 1 traffic vehicle model can perceive information about vehicles in the same lane and on the left and right adjacent lanes. Compared to the Level 0 traffic vehicle model, it makes more accurate and rapid longitudinal decisions and proactively determines whether vehicles in adjacent lanes are cutting in. It can only perform longitudinal driving control and has a preliminary understanding of the traffic environment.

[0018] The Level 2 traffic vehicle model can perceive information about vehicles in the same lane and on the left and right adjacent lanes. Compared to the Level 1 traffic vehicle model, it makes more accurate and rapid longitudinal decisions, proactively determines whether adjacent lane vehicles are cutting in, can proactively respond to adjacent vehicle behavior, and can perform lateral and longitudinal control, providing a preliminary understanding of the traffic environment.

[0019] The Level 3 traffic vehicle model can perceive information about vehicles in the same lane and the left and right lanes. Compared with the Level 2 traffic vehicle, it can achieve integrated lateral and longitudinal control, enhance the continuity and internal correlation of driving behavior, and can adaptively learn the optimal decision-making behavior. It has strong interaction with surrounding vehicles and has a high-level understanding of the traffic environment.

[0020] The specific method of the second step is as follows:

[0021] Step 1: For level 0 traffic vehicles, considering that the longitudinal model needs to deal with the problem of the preceding vehicle following, the acceleration of the vehicle is dynamically adjusted to maintain a safe and comfortable driving distance. The intelligent driver model (IDM model) is used as the level 0 traffic vehicle model, and its structure is shown in formula (1):

[0022]

[0023] The expected following distance of the host vehicle to avoid collision is:

[0024]

[0025] Where a max is the maximum acceleration of the main vehicle, δ is the acceleration index, v t is the current speed of the main vehicle, v e is the expected speed of the main vehicle, Δv t The relative speed between the main vehicle and the preceding vehicle, s t is the relative distance between the main vehicle and the preceding vehicle, b is the comfortable deceleration, T is the reaction time, s e For safe parking distance;

[0026] Step 2: For the Level 1 traffic vehicle model, considering that the longitudinal decision-making model needs to be able to dynamically adjust the vehicle's acceleration based on the speed and position information of the vehicle ahead to maintain a safe and comfortable driving distance; at the same time, it needs to adapt to more complex traffic scenarios and implement accurate and flexible decision-making strategies. A longitudinal decision-making model based on the IDM-LSTM combination is designed. The adaptive Kalman filter dynamically adjusts the linear combination weights of the IDM and LSTM algorithms, taking into account the safety and comfort of the IDM algorithm and the accuracy of the LSTM algorithm.

[0027] The laser radar and millimeter-wave radar visual sensors installed on the main vehicle collect data on the vehicles in front and behind the main vehicle, including the speed of the leading vehicle, the acceleration of the leading vehicle, the relative speed of the leading vehicle and the main vehicle, and the relative distance between the leading vehicle and the main vehicle;

[0028] Based on the above collected data, the IDM model is calibrated using the simulated annealing optimization parameter method, namely the acceleration index δ, reaction time T, and safe parking distance s. e , comfortable deceleration b, expected vehicle speed v e , maximum speed a max ;

[0029] Construct an LSTM vehicle following control offline prediction network. First, extract the front vehicle speed v from the previously collected driving data set. f , main vehicle speed v h , relative speed Δv, acceleration of the preceding vehicle a f , host vehicle acceleration a h And the relative distance Δx as input features, the expected acceleration a of the main vehicle exp and expected following distance v exp As output label;

[0030] Afterwards, the dataset is divided into training, validation, and test sets. The data is then split into several sequence samples according to the set time window. An LSTM network model is constructed, consisting of an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer receives the six features mentioned above, the LSTM layer is used to extract time series information, the fully connected layer performs dimensionality reduction and nonlinear transformation, and maps the multidimensional data output to two-dimensional data. The output layer outputs two predicted values.

[0031] The LSTM network model was trained and validated using mean squared error (MSE) and Adam as the loss function and optimizer, with hyperparameters and network structure adjusted until satisfactory performance was achieved. The predictions of the LSTM network model were evaluated using a test set, and the root mean squared error (RMSE) between the predicted and true values ​​was calculated. The trained LSTM network model was saved as an offline prediction network model and used as part of the vehicle longitudinal control model. In practical applications, the offline prediction network model was used to determine the expected following distance and expected acceleration of the driver vehicle using real-time and historical data collected from the vehicle.

[0032] In addition, a vector conversion method is designed to determine whether a vehicle in the left and right lanes will cut into the host lane based on the operating status of the vehicles in the left and right lanes, the main vehicle's operating speed, and the speed and direction of the vehicles in the side lanes, so that corresponding actions can be taken in advance.

[0033] If the front vehicle is in the process of changing lanes and cutting in, the relative distance between the host vehicle and the front vehicle should be recalculated. At the same time, the host vehicle will respond to the front vehicle cutting in in advance. First, it is necessary to determine whether the front vehicle is changing lanes or cutting in. By comparing the longitudinal distance x p and the horizontal distance y p The rate of change is used to judge and It is obtained by directly derivation of the distance. Considering that the front vehicle may change lanes in the direction away from the main vehicle, if y is detected p Increases over a period of time, then is 0, when If it is 0, it is considered that the vehicle is traveling in a straight line in different lanes or changing lanes in the direction of the main vehicle, and it will not be considered first. When it is not 0, it is considered that the vehicle is changing lanes close to the host vehicle and the relative distance needs to be calculated. The relative distance x c The formula is:

[0034]

[0035] in and The calculation formulas are:

[0036]

[0037]

[0038] The longitudinal decision model makes a corresponding decision based on the recalculated relative distance to the preceding vehicle's cutting-in behavior. This behavior reflects the vehicle's preliminary understanding of the traffic situation in the adjacent lane and its ability to respond accordingly.

[0039] Step 3: For the Level 2 traffic model, the longitudinal control decision is the same as that for the Level 1 traffic model. A rule-based finite state machine is used as the decision model for lateral lane changes. By presetting lane-changing rules, the interaction between the vehicle and surrounding vehicles is simulated. The process is as follows:

[0040] When there is a vehicle ahead of the main vehicle, the main vehicle enters the following state. If the following state is maintained for a set time and the desired speed is not reached, it switches to the lane change state. If the surrounding lanes meet better driving conditions, lane change preparation is carried out and the lane change conditions are judged.

[0041] Lane changing conditions take safety into consideration and use a TTC indicator based on a safe distance model. If there are vehicles in the target lane, and the main vehicle's speed is lower than that of the rear vehicle, or higher than that of the front vehicle, a TTC determination is made. If the TTC exceeds the threshold, safety conditions are met and the lane change is executed.

[0042] During the lane change process, the vehicle's yielding safety rules are designed to determine whether the following vehicle will cut in. If the following vehicle accelerates more than 1m / s during the lane change process, 2 , based on safety considerations, wait for the opportunity to change lanes again to improve driving safety;

[0043] Step 4: For the Level 3 traffic vehicle model, considering the vehicle's higher intelligence and in-depth understanding of traffic conditions, the deep reinforcement learning algorithm DQN is used as the decision-making control model to perform integrated control of longitudinal control and lane change decisions. Compared to the lateral and longitudinal decoupled control of the Level 2 model, the continuity and inherent connection of driving behavior are enhanced, and the interaction with surrounding vehicles is strengthened. Define the state space, action space, and reward function. The state space contains information on the position, velocity, acceleration, and heading angle of the vehicle and surrounding vehicles. The action space includes two dimensions: longitudinal acceleration and lateral lane change. The reward function comprehensively considers factors such as driving efficiency, safety, and comfort.

[0044] Construct a DQN network model, including a Q network and a target network. The Q network is used to estimate the Q value of each action in the current state, and the target network is used to estimate the maximum Q value of the next state. The structure of the two networks can be composed of multiple fully connected 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.

[0045] We selected mean squared error and Adam as the loss function and optimizer to train the DQN network model. During training, we used the ε-greedy strategy to explore and exploit the environment, the experience replay mechanism to store and sample transfer data, and a fixed frequency to update the parameters of the target network.

[0046] Use the test set to evaluate the control effect of the DQN network model and calculate the error indicators between the predicted actions and the actual actions, such as the root mean square error and mean absolute error;

[0047] The trained DQN network model is saved as an integrated control network model, which can be used in vehicle-following and lane-changing scenarios. In actual applications, the integrated control network model uses real-time and historical data collected by vehicle-to-vehicle communication technology to obtain the optimal longitudinal acceleration and lateral lane change.

[0048] The specific method of the third step is as follows:

[0049] Step 1: Discretize the road space into micro-grids of different sizes. The size of a single grid is Δx × Δy, where Δx and Δy are determined by the lane width and vehicle size.

[0050] First, the road space is divided into several strips of length Δx along the road direction. Then, each strip is divided into several small areas of width Δy along the road direction perpendicular to the road. In this way, the area of ​​each grid is Δx×Δy. Specifically, Δx is greater than or equal to the length of the vehicle, and Δy is greater than or equal to the width of the vehicle.

[0051] The purpose of grid division is to discretize the road space into several regions, each of which represents the position and status of a vehicle. The principle of grid division is to make the size of each region match the size of the vehicle, so as to avoid excessive number of grids, which would affect the efficiency of subsequent optimization.

[0052] Step 2: Define the objective function f to represent the distribution of vehicles of different levels:

[0053]

[0054] Where n is the number of traffic vehicle models, w ij is the weight between the i-th and j-th traffic vehicle models, d ij is the distance between the i-th and j-th traffic vehicle models;

[0055] w ij The value of depends on the type of driver:

[0056]

[0057] If the i-th and j-th traffic vehicles have the same level, i.e., l i =l j , then w ij = -1, avoid distributing vehicle models of the same level to adjacent areas;

[0058] If the i-th and j-th drivers are of different types, i.e., l i ≠l j , then w ij =1\1.2\1.4, the specific value depends on the difference in traffic vehicle levels, so as to increase the mixing degree of traffic vehicle models of different levels.

[0059] d ij The calculation formula is:

[0060]

[0061] Step 3: Use the particle swarm optimization algorithm to find the optimal traffic model distribution scheme. Initialize a group of random particles of N, each particle represents a possible traffic model distribution scheme, that is, the coordinate position of each traffic vehicle in the grid. In the two-dimensional search space, the algorithm parameters are:

[0062] X id =(x ix ,x iy ) (9)

[0063] V id =(v ix ,v iy ) (10)

[0064] P id,best =(p ix ,p iy ) (11)

[0065] P d,gbest =(p x,gbest ,p y,gbest ) (12)

[0066] Among them, X id is the position of the i-th particle, V id is the velocity of the i-th particle, and includes the distance and direction of particle movement, P id，best is the optimal position of the i-th particle, P d，best The optimal position searched for the group;

[0067] Step 4. Then update the particle speed and position:

[0068]

[0069]

[0070] Where k is the number of iterations, w is the inertia weight, c1 is the individual learning factor, c2 is the group learning factor, r1 and r2 are random numbers in the interval [0,1] to increase the randomness of the search;

[0071] Step 5. Repeat the above steps to calculate the objective function value f(x) of each particle and update the speed and position of each particle until the termination condition is met, that is, the number of iterations is greater than the set maximum number of iterations of 50 times, and output the global optimal solution g, which is the optimal traffic vehicle model distribution scheme, and then complete the construction of the heterogeneous mixed traffic flow model.

[0072] Beneficial effects of the present invention:

[0073] The novel hierarchical evaluation and construction method for heterogeneous mixed traffic flow models provided by this invention establishes a multidimensional hierarchical evaluation index for a unified hierarchical evaluation of autonomous vehicles in heterogeneous mixed traffic flows. This establishes a hierarchical index that better reflects traffic flow characteristics and achieves a uniform distribution of autonomous vehicles of different functions, i.e., different levels, on a given road, thereby constructing a heterogeneous mixed traffic flow simulation model. This method can be applied to provide an effective, high-confidence heterogeneous mixed traffic flow model for intelligent vehicle simulation research. It can also be extended to research in areas such as intelligent vehicle decision-making, planning, control, and testing, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 Schematic diagram of the hierarchical evaluation and construction process of the heterogeneous mixed traffic flow model described in the present invention.

[0075] Figure 2 This is a flow chart of the IDM-LSTM network based on adaptive Kalman filtering described in the present invention.

[0076] Figure 3 This is a schematic diagram of the side car cutting in according to the present invention.

[0077] Figure 4 This is a schematic diagram of the state machine structure for implementing the lateral lane-changing strategy described in the present invention.

[0078] Figure 5 This is a flow chart of the traffic vehicle model distribution based on micro-element grid weight distribution according to the present invention. DETAILED DESCRIPTION

[0079] See also Figures 1 to 5 As shown:

[0080] The present invention provides a novel hierarchical evaluation and construction method for heterogeneous mixed traffic flow models, which includes the following steps:

[0081] The first step is to establish a functional grading evaluation index for autonomous driving vehicles, and to uniformly grade and evaluate vehicles with different autonomous driving functions. The details are as follows:

[0082] The different functions of autonomous driving vehicles can be reflected in traffic flow as the degree and scope of the vehicles' ability to adapt and cooperate in complex traffic environments, which can be reflected in the interactive game between the vehicle and surrounding vehicles and the vehicle's ability to make autonomous decisions. Therefore, based on the interaction between autonomous driving vehicles and surrounding vehicles and their own decision-making and control capabilities, three grading criteria are proposed, namely perception range, control function, and depth of understanding.

[0083] Specifically, the perception range refers to the road area and traffic vehicle information that the vehicle can perceive; the control function refers to the control capability that the vehicle can make; and the depth of understanding refers to the vehicle's cognition of the surrounding vehicle behavior and road environment, as well as the vehicle's ability to actively respond to the surrounding vehicle behavior.

[0084] Based on the three aforementioned evaluation criteria, autonomous vehicles are divided into four levels, as shown in Table 1. As the level increases, the vehicle's interactivity with surrounding vehicles increases, and its ability to adapt to complex traffic situations, making reasonable and predictable real-world actions, and improving traffic efficiency increases.

[0085] Table 1 Classification of heterogeneous traffic flow vehicle models

[0086]

[0087] Specifically, the Level 0 traffic vehicle model can only perceive information about vehicles in front of it in the same lane and can only perform longitudinal driving control. It has no understanding of the traffic environment and can only respond passively.

[0088] The Level 1 traffic vehicle model can perceive information about vehicles in the same lane and the left and right lanes. Compared with the Level 0 traffic vehicle model, it makes longitudinal decisions more accurately and quickly, and can actively judge whether vehicles in the side lanes will cut in and interact. It can only perform longitudinal driving control and has a preliminary understanding of the traffic environment.

[0089] The Level 2 traffic vehicle model can perceive information about vehicles in the same lane and the left and right lanes. Compared with the Level 1 traffic vehicle model, it makes more accurate and rapid longitudinal decisions, actively judges whether vehicles in the adjacent lanes will cut in and interact, can actively respond to the behavior of adjacent vehicles, perform lateral and longitudinal control, and have a preliminary understanding of the traffic environment.

[0090] The Level 3 traffic vehicle model can perceive information about vehicles in the same lane and the left and right lanes. Compared with the Level 2 traffic vehicle, it can achieve integrated lateral and longitudinal control, enhance the continuity and internal correlation of driving behavior, adaptively learn the optimal decision-making behavior, have strong interaction with surrounding vehicles, and have a high-level understanding of the traffic environment.

[0091] The second step is to design decision-making control algorithms for each level of autonomous driving function according to the grading indicators, generating a multi-level autonomous driving traffic model to reflect the operation of vehicles with different autonomous driving functions.

[0092] For level 0 traffic vehicles, the longitudinal model needs to deal with the problem of the preceding vehicle following the preceding vehicle, dynamically adjust the vehicle's acceleration, and maintain a safe and comfortable driving distance. Since the intelligent driver model (IDM model) can well simulate the driver's following behavior with the preceding vehicle in different traffic flows and has good performance, it can achieve stable and smooth following. Therefore, the IDM model is used as the level 0 traffic vehicle model. Its structure is shown in formula (1):

[0093]

[0094] Furthermore, the formula can be divided into two parts: Describes the acceleration of a vehicle on a non-congested road. Indicates the interactive braking acceleration when the host vehicle approaches the vehicle in front.

[0095] The expected following distance of the host vehicle to avoid collision is:

[0096]

[0097] Where a max is the maximum acceleration of the main vehicle, δ is the acceleration index, v t is the current speed of the main vehicle, v e is the expected speed of the main vehicle, Δv t The relative speed between the main vehicle and the preceding vehicle, s t is the relative distance between the main vehicle and the preceding vehicle, b is the comfortable deceleration, T is the reaction time, s e For safe parking distance.

[0098] For the Level 1 traffic vehicle model, the longitudinal decision model needs to be able to dynamically adjust the vehicle's acceleration based on the speed and position information of the vehicle in front to maintain a safe and comfortable driving distance; at the same time, it needs to adapt to more complex traffic scenarios and implement accurate and flexible decision-making strategies. Figure 2 The adaptive Kalman filter dynamically adjusts the linear combination weights of the IDM and LSTM algorithms, balancing the safety and comfort of the IDM algorithm with the accuracy of the LSTM algorithm. Specifically, the state vector is the acceleration difference between the IDM and LSTM algorithm outputs, and the observation vector is the acceleration output of the IDM and LSTM. The adaptive Kalman filter optimally estimates the acceleration difference and then adds it to the LSTM predicted acceleration to obtain the final output acceleration, which is:

[0099]

[0100] in is the posterior estimate of the acceleration difference obtained by the adaptive Kalman filter.

[0101] The laser radar, millimeter-wave radar and other visual sensors installed on the main vehicle collect data on the vehicles in front and behind the main vehicle, including the speed of the leading vehicle, the acceleration of the leading vehicle, the relative speed of the leading vehicle and the main vehicle, and the relative distance between the leading vehicle and the main vehicle.

[0102] Based on the above collected data, the IDM model is calibrated using the simulated annealing optimization parameter method, namely the acceleration index δ, reaction time T, and safe parking distance s. e , comfortable deceleration b, expected vehicle speed v e , maximum speed a max First, the initial values ​​of the parameters to be calibrated are randomly generated. A new parameter solution is randomly generated within the range of the current random parameter solution. The objective function value of the parameter solution is calculated. If the value of the objective function of the new solution is less than the objective function value of the current parameter solution, the new solution is selected as the current solution. This process is then iterated until the optimal solution is obtained after the number of iterations is reached and the termination condition is met.

[0103] Construct an LSTM vehicle following control offline prediction network. First, extract the front vehicle speed v from the previously collected driving data set. f , main vehicle speed v h , relative speed Δv, acceleration of the preceding vehicle a f , host vehicle acceleration a h And the relative distance Δx as input features, the expected acceleration a of the main vehicle exp and expected following distance v exp as the output label.

[0104] Next, the dataset is divided into training, validation, and test sets, with the data segmented into multiple sequence samples according to specific time windows. An LSTM network model is constructed, consisting of an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer receives the six features described above, the LSTM layer extracts time series information, and the fully connected layer performs dimensionality reduction and nonlinear transformation, mapping the multidimensional data output to two-dimensional data. The output layer outputs two predicted values.

[0105] Specifically, [v f ,v h ,Δv,a f ,a h ,Δx] normalized as the input x of the model t , after LSTM layer, fully connected layer and output layer, after denormalization output [a exp ,x exp ].

[0106] The LSTM network model was trained and validated using mean squared error (MSE) and Adam as the loss function and optimizer, adjusting hyperparameters and network structure until satisfactory performance was achieved. The prediction performance of the LSTM network model was evaluated using a test set, calculating the root mean squared error (RMSE) between the predicted and true values. The trained LSTM network model was saved as an offline prediction network model and used as part of the vehicle longitudinal control model. In practical applications, the offline prediction network model was used to determine the expected following distance and expected acceleration of the host vehicle using real-time and historical data collected by the vehicle.

[0107] In addition, a vector conversion method is designed to determine whether a vehicle in the left and right lanes will cut into the lane based on the vehicle's speed and the speed and direction of the vehicle, so that corresponding actions can be taken in advance. Figure 3 .

[0108] Specifically, when the front car is changing lanes and cutting in, the relative distance between the host car and the front car should be recalculated, and the host car will respond to the front car cutting in in advance. First, it is necessary to determine whether the front car is changing lanes or cutting in by comparing the longitudinal distance x p and the horizontal distance y p The rate of change is used to judge and It is obtained by directly derivation of the distance. Considering that the front vehicle may change lanes in the direction away from the main vehicle, if y is detected p Increases over a period of time, then is 0. If it is 0, it can be considered that the vehicle is driving in a straight line in different lanes or changing lanes in the direction of the main vehicle, and it will not be considered. When it is not 0, it can be considered that the vehicle is changing lanes close to the host vehicle, and the relative distance needs to be calculated. The relative distance x c The formula is:

[0109]

[0110] in and The calculation formulas are:

[0111]

[0112]

[0113] The longitudinal decision model makes corresponding decisions based on the recalculated relative distance in response to the preceding vehicle's cutting-in behavior. This behavior reflects that the main vehicle has a preliminary understanding of the traffic situation in the adjacent lane and can respond based on the behavior of the adjacent vehicle.

[0114] For the Level 2 traffic vehicle model, the longitudinal control decision is the same as that for the Level 1 traffic vehicle model. A rule-based finite state machine is used as the decision model for lateral lane change. By presetting lane change rules, the interaction process between the vehicle and surrounding vehicles is simulated. Figure 4 The process is:

[0115] When a leading vehicle is in front of the host vehicle, the host vehicle enters the following state. During the following process, if the leading vehicle moves away, the host vehicle maintains the cruising state. If the following state is maintained for a certain period of time and the desired speed is not reached, the host vehicle generates a lane change intention and switches to the lane change intention state. If the surrounding lanes meet better driving conditions, the host vehicle begins to prepare for the lane change and begins to determine the lane change conditions. If the conditions are met, the host vehicle changes lanes. If not, the host vehicle continues to follow the leading vehicle.

[0116] The lane-changing condition takes safety factors into consideration and selects the TTC indicator based on the safety distance model. If there are vehicles in the target lane, a TTC judgment is made when the vehicle's speed is lower than that of the vehicle behind it, or when the vehicle's speed is higher than that of the vehicle in front of it. If the TTC is greater than the threshold, the safety condition is met and the lane-changing operation is performed.

[0117] During the lane change process, the vehicle's yielding safety rules are designed to determine whether the following vehicle will cut in. If the following vehicle accelerates more than 1m / s during the lane change process, 2 , then based on safety considerations, wait for the opportunity to change lanes again to improve driving safety.

[0118] For the Level 3 traffic vehicle model, considering the vehicle's higher intelligence and in-depth understanding of traffic conditions, the deep reinforcement learning algorithm DQN is used as the decision-making and control model for integrated longitudinal control and lane change decisions. Compared to the decoupled lateral and longitudinal control of the Level 2 model, this enhances the continuity and inherent connectivity of driving behavior, as well as interaction with surrounding vehicles. The state space, action space, and reward function are defined. The state space includes information such as the position, velocity, acceleration, and heading angle of the ego vehicle and surrounding vehicles. The action space includes longitudinal acceleration and lateral lane change dimensions. The reward function comprehensively considers factors such as driving efficiency, safety, and comfort.

[0119] Build a DQN network model, consisting of a Q network and a target network. The Q network estimates the Q value of each action in the current state, while the target network estimates the maximum Q value for the next state. The structure of these two networks can consist of multiple fully connected layers, with the input layer dimensioned by the size of the state space and the output layer dimensioned by the size of the action space.

[0120] The DQN network model was trained using mean squared error and Adam as the loss function and optimizer. During training, an ε-greedy strategy was used to explore and exploit the environment, an experience replay mechanism was used to store and sample transfer data, and a fixed frequency was used to update the target network parameters.

[0121] Use the test set to evaluate the control effect of the DQN network model and calculate the error indicators between the predicted actions and the actual actions, such as the root mean square error and mean absolute error.

[0122] The trained DQN network model is saved as an integrated control network model, which can be used in car-following and lane-changing scenarios. In practical applications, the integrated control network model uses real-time and historical data collected through vehicle-to-vehicle communication technology to achieve optimal longitudinal acceleration and lateral lane changes.

[0123] The third step is to construct a mixed traffic flow model, design a driving model distribution algorithm based on micro-element grid weight allocation, and evenly distribute the driving models described in step 2 with different set proportions on the road to construct a heterogeneous mixed traffic flow model. Figure 5 .

[0124] First, the road space is discretized into micro-element grids of different sizes. The size of a single grid is Δx×Δy, where Δx and Δy are determined by the road lane width and the vehicle size.

[0125] First, divide the road space into several strips of length Δx along the road direction. Then, divide each strip into several smaller areas of width Δy perpendicular to the road direction, i.e., grids. Thus, the area of ​​each grid is Δx × Δy. Specifically, Δx should be greater than or equal to the vehicle length, and Δy should be greater than or equal to the vehicle width.

[0126] The purpose of grid division is to discretize the road space into several areas, each of which represents the position and status of a traffic vehicle. The principle of grid division is to make the size of each area match the size of the vehicle, which can avoid excessive number of grids and affect the efficiency of subsequent optimization.

[0127] Then, define the objective function f to represent the distribution of vehicles of different levels:

[0128]

[0129] Where n is the number of traffic vehicle models, w ij is the weight between the i-th and j-th traffic vehicle models, d ij is the distance between the i-th and j-th traffic vehicle models.

[0130] w ijThe value of depends on the type of driver:

[0131]

[0132] If the i-th and j-th traffic vehicles have the same level, i.e., l i =l j , then w ij =-1, to prevent the same level of traffic vehicle models from being distributed to adjacent areas.

[0133] If the i-th and j-th drivers are of different types, i.e., l i ≠l j , then w ij =1\1.2\1.4, the specific value depends on the difference in traffic vehicle levels, so as to increase the mixing degree of traffic vehicle models of different levels.

[0134] d ij The calculation formula is:

[0135]

[0136] The particle swarm optimization algorithm is used to find the optimal traffic model distribution scheme. A group of random particles N is initialized. Each particle represents a possible traffic model distribution scheme, that is, the coordinate position of each traffic vehicle in the grid. In the two-dimensional search space, the algorithm parameters are:

[0137] X id =(x ix ,x iy ) (twenty three)

[0138] V id =(v ix ,v iy ) (twenty four)

[0139] P id,best =(p ix ,p iy ) (25)

[0140] P d,gbest =(p x,gbest ,p y,gbest ) (26)

[0141] Among them, X id is the position of the i-th particle, V id is the velocity of the ith particle (including the distance and direction of particle movement), P id，best is the optimal position of the i-th particle, P d，best The optimal position searched for the group.

[0142] The particle velocity and position are then updated:

[0143]

[0144]

[0145] Specifically, the velocity update consists of three components: inertia, cognition, and society. This can be interpreted as the particle's next iteration's movement direction = inertia direction + individual optimal direction + group optimal direction. Here, k is the number of iterations, w is the inertia weight, c1 is the individual learning factor, c2 is the group learning factor, and r1 and r2 are random numbers in the interval [0, 1] that increase the randomness of the search.

[0146] Finally, the above steps are repeated repeatedly, calculating the objective function value f(x) for each particle and updating its velocity and position until the termination condition is met (i.e., the number of iterations exceeds the set maximum number of iterations). The global optimal solution g is output, which is the optimal traffic model distribution solution, thus completing the construction of the heterogeneous mixed traffic flow model.

[0147] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any appropriate manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A novel hierarchical evaluation and construction method for heterogeneous mixed traffic flow models, characterized by: The method includes the following steps: The first step is to establish a functional grading evaluation index for autonomous driving vehicles, and conduct a unified grading evaluation for vehicles with different autonomous driving functions; The second step is to design decision-making control algorithms for each level of autonomous driving function according to the grading indicators, generating a multi-level autonomous driving traffic model to reflect the operation of vehicles with different autonomous driving functions. The third step is to construct a mixed traffic flow model. We design a driving model distribution algorithm based on micro-grid weight allocation, distribute the traffic vehicle models described in step 2 in different proportions evenly on the road, and construct a heterogeneous mixed traffic flow model. The specific method is as follows: Step 1: Discretize the road space into micro-grids of different sizes. The size of a single grid is Δx × Δy, where Δx and Δy are determined by the lane width and vehicle size. First, the road space is divided into several strips of length Δx along the road direction. Then, each strip is divided into several small areas of width Δy along the road direction perpendicular to the road. In this way, the area of ​​each grid is Δx×Δy. Specifically, Δx is greater than or equal to the length of the vehicle, and Δy is greater than or equal to the width of the vehicle. The purpose of grid division is to discretize the road space into several regions, each of which represents the position and status of a vehicle. The principle of grid division is to make the size of each region match the size of the vehicle, so as to avoid excessive number of grids, which would affect the efficiency of subsequent optimization. Step 2: Define the objective function f to represent the distribution of vehicles of different levels: Where n is the number of traffic vehicle models, w ij is the weight between the i-th and j-th traffic vehicle models, d ij is the distance between the i-th and j-th traffic vehicle models; w ij The value of depends on the type of driver: If the i-th and j-th traffic vehicles have the same level, i.e., l i =l j , then w ij = -1, avoid distributing vehicle models of the same level to adjacent areas; If the i-th and j-th drivers are of different types, i.e., l i ≠l j , then w ij =1\1.2\1.4, the specific value depends on the difference in traffic vehicle levels, so as to increase the mixing degree of traffic vehicle models of different levels; d ij The calculation formula is: Step 3: Use the particle swarm optimization algorithm to find the optimal traffic model distribution scheme. Initialize a group of random particles of N, each particle represents a possible traffic model distribution scheme, that is, the coordinate position of each traffic vehicle in the grid. In the two-dimensional search space, the algorithm parameters are: 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) Among them, X id is the position of the i-th particle, V id is the velocity of the i-th particle, and includes the distance and direction of particle movement, P id，best is the optimal position of the i-th particle, P d，best The optimal position searched for the group; Step 4. Then update the particle speed and position: Where k is the number of iterations, w is the inertia weight, c1 is the individual learning factor, c2 is the group learning factor, r1 and r2 are random numbers in the interval [0,1] to increase the randomness of the search; Step 5. Repeat the above steps to calculate the objective function value f(x) of each particle and update the speed and position of each particle until the termination condition is met, that is, the number of iterations is greater than the set maximum number of iterations of 50 times, and output the global optimal solution g, which is the optimal traffic vehicle model distribution scheme, and then complete the construction of the heterogeneous mixed traffic flow model.

2. A novel hierarchical evaluation and construction method for heterogeneous mixed traffic flow models according to claim 1, characterized in that: The specific method of the first step is as follows: Based on the interaction between autonomous vehicles and surrounding vehicles and their own decision-making and control capabilities, three classification criteria are proposed: perception range, control function, and depth of understanding. The details are as follows: The perception range refers to the road area and traffic information that the vehicle can perceive; the control function refers to the control capability of the vehicle; and the understanding depth refers to the vehicle's cognition of the surrounding vehicle behavior and road environment, as well as the vehicle's ability to actively respond to the surrounding vehicle behavior. Based on the three aforementioned evaluation criteria, autonomous vehicles are divided into four levels. From low to high, the interactivity of the vehicle with surrounding vehicles increases successively. At the same time, it can adapt to complex traffic games, make reasonable and predictable real actions, and improve traffic efficiency. The classification of heterogeneous traffic flow vehicle models is as follows: The Level 0 traffic vehicle model can only perceive information about vehicles in the same lane ahead and can only perform longitudinal driving control. It has no understanding of the traffic environment and can only react passively. The Level 1 traffic vehicle model can perceive information about vehicles in the same lane and on the left and right adjacent lanes. Compared to the Level 0 traffic vehicle model, it makes more accurate and rapid longitudinal decisions and proactively determines whether vehicles in adjacent lanes are cutting in. It can only perform longitudinal driving control and has a preliminary understanding of the traffic environment. The Level 2 traffic vehicle model can perceive information about vehicles in the same lane and on the left and right adjacent lanes. Compared to the Level 1 traffic vehicle model, it makes more accurate and rapid longitudinal decisions, proactively determines whether adjacent lane vehicles are cutting in, can proactively respond to adjacent vehicle behavior, and can perform lateral and longitudinal control, providing a preliminary understanding of the traffic environment. The Level 3 traffic vehicle model can perceive information about vehicles in the same lane and the left and right lanes. Compared with the Level 2 traffic vehicle, it can achieve integrated lateral and longitudinal control, enhance the continuity and internal correlation of driving behavior, and can adaptively learn the optimal decision-making behavior. It has strong interaction with surrounding vehicles and has a high-level understanding of the traffic environment.

3. The novel hierarchical evaluation and construction method of heterogeneous mixed traffic flow model according to claim 1 is characterized by: The specific method of the second step is as follows: Step 1: For level 0 traffic vehicles, considering that the longitudinal model needs to deal with the problem of the preceding vehicle following, the acceleration of the ego vehicle is dynamically adjusted to maintain a safe and comfortable driving distance. The intelligent driver model is used as the level 0 traffic vehicle model, and its structure is shown in formula (1): The expected following distance of the host vehicle to avoid collision is: Where a max is the maximum acceleration of the main vehicle, δ is the acceleration index, v t is the current speed of the main vehicle, v e The expected speed of the main vehicle, Δvt is the relative speed between the main vehicle and the preceding vehicle, s t is the relative distance between the main vehicle and the preceding vehicle, b is the comfortable deceleration, T is the reaction time, s e For safe parking distance; Step 2: For the Level 1 traffic vehicle model, considering that the longitudinal decision-making model needs to be able to dynamically adjust the vehicle's acceleration based on the speed and position information of the vehicle ahead to maintain a safe and comfortable driving distance; at the same time, it needs to adapt to more complex traffic scenarios and implement accurate and flexible decision-making strategies. A longitudinal decision-making model based on the IDM-LSTM combination is designed. The adaptive Kalman filter dynamically adjusts the linear combination weights of the IDM and LSTM algorithms, taking into account the safety and comfort of the IDM algorithm and the accuracy of the LSTM algorithm. The laser radar and millimeter-wave radar visual sensors installed on the main vehicle collect data on the vehicles in front and behind the main vehicle, including the speed of the leading vehicle, the acceleration of the leading vehicle, the relative speed of the leading vehicle and the main vehicle, and the relative distance between the leading vehicle and the main vehicle; Based on the collected data, the intelligent driver model is calibrated using the simulated annealing optimization parameter method, namely the acceleration index δ, reaction time T, and safe parking distance s. e , comfortable deceleration b, expected vehicle speed v e , maximum speed a max ; Construct an LSTM vehicle following control offline prediction network. First, extract the front vehicle speed v from the collected driving data set. f , main vehicle speed v h , relative speed Δv, acceleration of the preceding vehicle a f , host vehicle acceleration a h And the relative distance Δx as input features, the expected acceleration a of the main vehicle exp and expected following distance v exp As output label; Afterwards, the dataset is divided into training, validation, and test sets. The data is then split into several sequence samples according to the set time window. An LSTM network model is constructed, consisting of an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer receives the aforementioned input features, the LSTM layer is used to extract time series information, the fully connected layer performs dimensionality reduction and nonlinear transformation, and maps the multidimensional data output to two-dimensional data. The output layer outputs two predicted values. The LSTM network model was trained and validated using mean squared error (MSE) and Adam as the loss function and optimizer, with hyperparameters and network structure adjusted until satisfactory performance was achieved. The predictions of the LSTM network model were evaluated using a test set, and the root mean squared error (RMSE) between the predicted and true values ​​was calculated. The trained LSTM network model was saved as an offline prediction network model and used as part of the vehicle longitudinal control model. In practical applications, the offline prediction network model was used to determine the expected following distance and expected acceleration of the driver vehicle using real-time and historical data collected from the vehicle. In addition, a vector conversion method is designed to determine whether a vehicle in the left and right lanes will cut into the host lane based on the vehicle's speed and the speed and direction of the adjacent vehicle, so that corresponding actions can be taken in advance. If the front vehicle is in the process of changing lanes and cutting in, the relative distance between the host vehicle and the front vehicle should be recalculated. At the same time, the host vehicle will respond to the front vehicle cutting in in advance. First, it is necessary to determine whether the front vehicle is changing lanes or cutting in. By comparing the longitudinal distance x p and the horizontal distance y p The rate of change is used to judge and It is obtained by directly derivation of the distance. Considering that the front vehicle may change lanes in the direction away from the main vehicle, if y is detected p Increases over a period of time, then is 0, when If it is 0, it is considered that the vehicle is traveling in a straight line in different lanes or changing lanes in the direction of the main vehicle, and it will not be considered first. When it is not 0, it is considered that the vehicle is changing lanes close to the host vehicle and the relative distance needs to be calculated. The relative distance x c The formula is: in and The calculation formulas are: The longitudinal decision model makes a corresponding decision based on the recalculated relative distance to the preceding vehicle's cutting-in behavior. This behavior reflects the vehicle's preliminary understanding of the traffic situation in the adjacent lane and its ability to respond accordingly. Step 3: For the Level 2 traffic model, the longitudinal control decision is the same as that for the Level 1 traffic model. A rule-based finite state machine is used as the decision model for lateral lane changes. By presetting lane-changing rules, the interaction between the vehicle and surrounding vehicles is simulated. The process is as follows: When there is a vehicle ahead of it, the host vehicle enters the following state. If the following state is maintained for a set time and the desired speed is not reached, it switches to the lane change state. If the surrounding lanes meet better driving conditions, lane change preparation is carried out and the lane change conditions are judged. Lane changing conditions take safety factors into consideration and use a TTC indicator based on a safe distance model. If there are vehicles in the target lane, and the vehicle's speed is lower than that of the vehicle behind it, or higher than that of the vehicle ahead, a TTC judgment is made. If the TTC exceeds the threshold, safety conditions are met and the lane change is executed. During the lane change process, the vehicle's yielding safety rules are designed to determine whether the following vehicle will cut in. If the following vehicle accelerates more than 1m / s during the lane change process, 2 , based on safety considerations, wait for the opportunity to change lanes again to improve driving safety; Step 4: For the Level 3 traffic vehicle model, considering that the vehicle has higher intelligence and can have a deep understanding of traffic situations, the deep reinforcement learning algorithm DQN is used as the decision-making control model to perform integrated control of longitudinal control and lane change decisions. Compared with the lateral and longitudinal decoupling control of the Level 2 model, the continuity and intrinsic connection of driving behavior are enhanced, and the interaction with surrounding vehicles is enhanced. The state space, action space and reward function are defined. The state space contains information on the position, velocity, acceleration, and heading angle of the vehicle and surrounding vehicles. The action space includes two dimensions: longitudinal acceleration and lateral lane change. The reward function comprehensively considers factors such as driving efficiency, safety, and comfort. Construct a DQN network model, including a Q network and a target network. The Q network is used to estimate the Q value of each action in the current state, and the target network is used to estimate the maximum Q value of the next state. The structure of the two networks can be composed of multiple fully connected 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. We selected mean squared error and Adam as the loss function and optimizer to train the DQN network model. During training, we used the ε-greedy strategy to explore and exploit the environment, the experience replay mechanism to store and sample transfer data, and a fixed frequency to update the parameters of the target network. Use the test set to evaluate the control effect of the DQN network model and calculate the error index between the predicted action and the actual action. The error index is the root mean square error or mean absolute error. The trained DQN network model is saved as an integrated control network model, which can be used in vehicle-following and lane-changing scenarios. In actual applications, the integrated control network model uses real-time and historical data collected by vehicle-to-vehicle communication technology to obtain the optimal longitudinal acceleration and lateral lane change.

Images