Intelligent driving behavior decision-making method and device fusing complex network theory and partially observable markov decision process

Abstract

The application discloses a complex network theory and a part observable Markov decision process intelligent driving behavior decision method and equipment, by introducing a part observable Markov decision process (POMDP), an ideal decision model for multi-agent interaction in an uncertain environment is provided, with the increase of model dimension, the calculation complexity is also greatly increased, in order to solve the calculation complexity problem, the application introduces a complex network cognitive theory based on driving environment modeling, is used to evaluate important nodes and extract risk trees, to reduce the confidence space of the POMDP model, accelerate the model solution, so as to realize the real-time behavior decision in the highly interactive dynamic uncertain driving environment, compared with the behavior decision method using the finite state machine and RL, the application can be effectively operated in the highly interactive dynamic uncertain driving environment while ensuring the algorithm interpretability, and provides a new solution for the development of high-level intelligent driving system.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of decision planning of intelligent driving system, and particularly relates to an intelligent driving behavior decision method and device for highly interactive dynamic uncertain driving environment by combining a complex network cognitive theory and a partially observable Markov decision process (POMDP). BACKGROUND

[0002] As one of the core modules of intelligent driving technology, decision planning is the brain of intelligent driving system by receiving and processing various types of perception information to make behavior decision and trajectory planning. The classical intelligent driving behavior decision method generally adopts a finite state machine method based on rule design, which has good interpretability and is currently widely used in automatic auxiliary navigation driving systems of L2 intelligent driving vehicles. However, as the driving environment faced by intelligent driving vehicles becomes more complex, the classical decision planning method cannot meet the highly interactive dynamic uncertain driving environment, and the intelligent driving behavior decision method based on RL (reinforcement learning) generally assumes that the driving environment is a completely observable MDP process, which is almost difficult to achieve in the highly interactive dynamic uncertain driving environment, and thus is difficult to apply to real driving scenarios. SUMMARY

[0003] To solve the above problems, the application introduces a partially observable Markov decision process (POMDP), which provides an ideal mathematical decision model for multi-agent interaction in uncertain environment, but the computational complexity will greatly increase with the increase of model dimension. To solve the problem of computational complexity, the application introduces a complex network cognitive theory based driving environment modeling for evaluating important nodes and extracting risk trees to reduce the confidence space of POMDP model and accelerate model solving, so as to realize real-time behavior decision in highly interactive dynamic uncertain driving environment.

[0004] The application aims to provide an intelligent driving behavior decision method based on the fusion of complex network theory and partially observable Markov decision process, which innovatively designs a behavior decision method that can effectively operate in highly interactive dynamic uncertain driving environment.

[0005] To achieve the above purpose, the decision method provided by the application adopts the following technical scheme, including the following steps:

[0006] Step 1: applying a complex network based cognitive theory to dynamically model the driving environment, including:

[0007] Step 1.1: first, based on the complex network theory, regarding the ego vehicle and other traffic participants as network nodes, a dynamic complex network model is constructed:

[0008] G t = (P, E, W, Θ) t where G t is a dynamic complex network model; P = {p1, p2,..., pN } is a set of nodes in the network, p i is a node in the network, and N is the number of nodes in the network; E = {e1, e2,..., e K} is a set of edges of nodes in the network, e i is an edge of nodes in the network, and K is the number of edges of nodes in the network; W = {w1, w2,..., w k} is a set of weights of edges; and Θ is a movable area of nodes.

[0009] Θ is modeled as a smooth bounded surface:

[0010]

[0011] where Ω is the boundary of the smooth bounded surface.

[0012] Step 1.2: Dynamically model the nodes based on a two-degree-of-freedom vehicle model and a moving point model.

[0013] The two-degree-of-freedom vehicle model is:

[0014]

[0015] where the state variables of the model are the yaw rate r and the lateral velocity U y ; δ is the front wheel steering angle; F yf and F yr are the front and rear lateral forces generated in driving, respectively; a and b are the front and rear axles, respectively; m is the vehicle mass; I z is the yaw moment of inertia; and the longitudinal velocity U x is considered as a time-varying parameter. When the tire cornering characteristics are within the linear range, the model can be represented as:

[0016]

[0017] where C f and C r are the front and rear cornering stiffnesses, respectively.

[0018] The relevant nodes are considered as moving particles with a center of gravity, and a moving point model is established relative to the desired path:

[0019]

[0020] where Δψ and e are the heading angle deviation and lateral path deviation, respectively; s is the distance along the desired path; and k is the curvature of the desired path.

[0021] Step 1.3: A variable Gaussian safety field model based on risk center shift is proposed according to the field theory. The static safety field S is described by a two-dimensional Gaussian function, and the static safety field S sta can be expressed as:

[0022]

[0023] where x0 and y0 are the coordinates of the risk center (i.e., the coordinate origin centered on the ego vehicle); C a is the field strength coefficient; a x and b y are the radii of the major and minor axes of the ellipse (the ellipse is a magnification of the inscribed ellipse of the vehicle's bounding box, which can be equivalently represented by the aspect ratio of the vehicle).

[0024] When the vehicle moves, the risk center will shift, and the new risk center is O'(x'0, y'0):

[0025]

[0026] where v is the velocity vector; k v is the adjustment factor, and 0 < k v < 1 (forward) or -1 < k v < 0 (backward); β is the included angle with the x-axis, and a' x and b' y represent the radii of the major and minor axes of the ellipse after the risk center shifts. The dynamic safety field S dyn can be expressed as:

[0027]

[0028] Step 2: According to the sensitivity of human drivers to distance and reaction time, the space around the node is divided into the first cognitive domain, the second cognitive domain, and the out-of-domain space.

[0029] The range of the first cognitive domain is:

[0030] maxa' x ≤ s th1

[0031] s th1 = t c1 · v e

[0032] where s th1 is the first threshold; t c1The first cognitive reaction time of a human driver; v e This represents the maximum approach speed of other nodes in the environment.

[0033] The scope of the second cognitive domain is:

[0034] s th1 <maxa′ x ≤s th2

[0035] s th2 =t c2 ·v e

[0036] Among them, s th2 The second threshold; t c2 This refers to the second cognitive reaction time of a human driver. The space outside the second cognitive domain is defined as the extra-domain space.

[0037] Within the framework of a variable Gaussian safety field, a risk perception function is established between nodes:

[0038]

[0039] in, It is node p i At node p j Field strength at the location, It is node p j scalar velocity, direction angle θ i，j It is the velocity vector of node px. With node p i field intensity vector The included angle (clockwise is positive), k c This is the risk perception adjustment coefficient.

[0040] Step 3: Based on the established model, generate a complex network and evaluate important nodes and extract risk trees.

[0041] Step 3.1: The evolution process of complex networks, including:

[0042] 1) Set the vehicle node as the master node, connect the master node and other nodes in the first cognitive domain, calculate the weight of the corresponding connection, and sort them according to the weight. For ease of calculation, if there are multiple nodes in the connection direction with the master node, take the nearest node as the node in the first cognitive domain, and regard other nodes that are far away as nodes in the second cognitive domain. Do the same for the nodes in the second cognitive domain.

[0043] 2) In the first cognitive domain, select environmental nodes in sequence according to the weight sequence determined in the previous step, calculate their weights with other nodes and sort them, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0044] 3) Select the environmental nodes in the second cognitive domain and the nodes in the first cognitive domain, and connect the node pairs with the largest weights.

[0045] 4) In the second cognitive domain, select environmental nodes in sequence, calculate their weights relative to other nodes, sort them, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0046] Step 3.2: Evaluate important nodes and generate a risk tree, including:

[0047] 1) Let node p i The set of all neighboring nodes is N. i ; Node strength r i For node p i The sum of the weights of its neighboring nodes is then:

[0048]

[0049] 2) The average strength of all nodes in the network is defined as the network strength. <r>Then we have:

[0050]

[0051] 3) The node strength r i The ratio of the sum of the intensities of all nodes is defined as To assess the importance of a node, we have:

[0052]

[0053] Will The largest node is defined as an important node. Important nodes in the first cognitive domain are defined as first-class important nodes (nodes that are not main nodes). Important nodes in the second cognitive domain are defined as second-class important nodes.

[0054] 4) Define the tree generated by the master node, the first type of important nodes, and the second type of important nodes as a risk tree.

[0055] The above steps enable the construction of a cognitive model based on complex networks, while also evaluating important nodes in the network and extracting a risk tree.

[0056] Step 4: This invention applies POMDP theory to model the vehicle decision-making process, eliminating uncertainties in the driving environment.

[0057] Step 4.1: Let ε t This represents the driving environment centered on the vehicle at time t, including the cognitive model built in the previous steps, traffic signals, and grid map of static obstacles. Let z represent the state of vehicle i∈V at time t. When i=0, it represents the vehicle itself. If the state variable lacks a superscript, it represents all vehicles; if it lacks a subscript, it represents all time points. The observation value received by the vehicle itself is denoted as z. t And use the observations to evaluate the true state quantity x t The input to the behavioral decision-making module is denoted as... <z t , ε t The output is a set of continuous discrete state parameters, denoted as... Where H represents the time length of a planning period (in this invention, the planning period is set to 5 seconds), x t+i This represents the state of all vehicles at time t+i.

[0058] Step 4.2: The POMDP model is defined as a set of six elements.<X,A,Z,T,O,R> Where X, A, and Z represent the state space, behavior space, and observation space, respectively; T represents the behavior a taken by the vehicle at time t. t After ∈A, the state probability transition function T(x) at consecutive times t-1 a t x t ) = p(x t |x t-1 , a t ), O is the probability function between the state quantity and the observation quantity O(x t , z t ) = p(z t |x t ), by expressing the state quantity and the observation value in the form of the probability distribution, to reflect the uncertainty in the driving environment. R(x t-1 , a t ) is a reward function, indicating the real-time reward obtained by the ego vehicle after taking the action a t-1 ∈X in the state x t ∈A.

[0059] When the ego vehicle is in a real driving scene, the state space can be represented as where N is the number of traffic participants, t is the current time, the state quantity of the vehicle includes the pose information, the node strength constructed in the complex network model, and the hidden state of the vehicle (such as the behavior intention) that cannot be directly observed.

[0060] Then T can be represented as The ego vehicle is distinguished from other traffic participants, and it is assumed that the transformation of the state is instantaneous, so T can be further transformed into:

[0061]

[0062] wherein, is the assumed driver model, reflecting the upper-level decision-making process of other traffic participants; is the state probability transition model of other traffic participants, reflecting the kinematics process of other traffic participants executing the decision-making, and for the purpose of simplifying the solution of the model, the present application sets it as a motion point model.

[0063] In order to eliminate the uncertainty, the POMDP model makes decisions on the belief space , which is the probability distribution of the state X. The belief degree of the belief space will be updated from the initial belief b t after the traffic participant takes the action a t and receives the observation value z t-1 . The present application uses the Bayesian formula to update the belief degree:

[0064]

[0065] wherein η is a normalization factor. The observation value of the other vehicle can be obtained by the perception and prediction module of the ego vehicle, so there is:

[0066]

[0067] Assuming the observation processes are independent of each other, the belief update process can be further written as:

[0068]

[0069] As can be seen from the above formula, although the state transition process of each traffic participant is independent of each other, and all state quantities and observation values of all traffic participants are used, so that the belief update process is an interactive process, so that the hidden state of other traffic participants can be updated step by step in the belief update process.

[0070] Therefore, in the POMDP model, the purpose of decision-making is to find a best decision strategy π on the belief space * such that the ego vehicle obtains the maximum reward expectation after taking a series of discrete behaviors in the planning period, that is:

[0071]

[0072] where t0 is the starting time of the current planning; γ is the discount coefficient of the reward R, and 0≤γ≤1; is the initial belief; b t is the belief at time t. Starting from the starting time and the initial belief, the action space A and the observation space O are expanded step by step in the planning period t H , and the nodes of the belief tree are generated step by step in the belief space. By using the Bellman equation for each node, the optimal strategy can be found:

[0073]

[0074] where, is the optimal utility function of the belief; Q * (b, a) is the optimal value that can be obtained after taking a certain behavior in the belief space. Then the behavior and observation value corresponding to the optimal strategy of each node can be found by the belief tree:

[0075]

[0076]

[0077] Taking the above steps can obtain the state quantity corresponding to the optimal strategy of each node, including the belief the behavior and the observation value that is, for let The final decision strategy is obtained

[0078] Step 5: According to the model built above, the behavior decision of the ego vehicle is made.

[0079] Step 5.1: For the convenience of problem solving, the present application abstracts the high-level semantic-level behavior (such as straight, left turn, right turn, etc.) from the model of the traffic participants Replace the complex behavior action Therefore, in the decision-making process of the whole model, only the semantic-level behavior input by the ego vehicle is needed to complete the behavior decision of the ego vehicle.

[0080] The behavior sequence of the ego vehicle is defined as a set of ordered decision trees generated in time sequence, and the nodes of the tree are the pre-defined semantic-level behaviors generated in time steps within the planning period. The decision goal of the ego vehicle is to find a set of semantic-level behaviors starting from the root node within the whole planning period, including a series of nodes, taking the leaf node of the last stage as the root node of the next stage, and so on, so that the whole decision-making behavior space Φ can be defined by a limited number of semantic-level strategies.

[0081] The size of the decision space will increase exponentially with the depth of the decision tree. In order to reduce the computational complexity, the present application prunes the decision tree. In real scenarios, the driver generally does not frequently change the semantic-level behavior within the same planning period, so it is set that each period contains at most one operation of changing the semantic behavior. At the same time, the cognitive model based on complex network established in step 1 is used for further pruning of the decision tree. When the risk weight between the ego vehicle and the direction of a node in the first cognitive domain is greater than the set threshold, the lane changing operation in the corresponding direction is prohibited. By changing the size of the threshold, different driving styles of the ego vehicle can be reflected.

[0082] Step 5.2: The trajectory of other traffic participants observed by the perception module is recorded as the observation value z t , assuming that the observation value is noiseless and the state of the ego vehicle is known, the observation model can further evaluate part of the hidden intention of other traffic participants, and the hidden intention of other traffic participants can be mapped to the corresponding semantic-level behavior, i.e. is deterministic, the random behavior of other traffic participants can be further converted into a confidence state, and the multiple trajectories of other traffic participants can be predicted by the prediction module, which is recorded as the initial confidence of the confidence space The confidence is updated by the confidence update formula established in step 4, thereby completing the construction of the observation space Z and the confidence space .

[0083] Step 5.3: Define the semantic-level behaviors of the ego vehicle and other traffic participants as a scene for closed-loop forward simulation, and the probability of each scene can be obtained through the joint distribution of the confidence probability of each traffic participant, so the number of scenes for simulation will exponentially increase with the number of traffic participants, which will make the complexity of the algorithm difficult to calculate, but in the actual driving scene, we often do not need to pay attention to all vehicles.

[0084] To solve this problem, the application uses the constructed cognitive model to prune the number of simulated scenes, sets the evaluation selection range to the nodes in the second cognitive domain (maintains straight driving when there are no other traffic participants in the second cognitive domain), and combines the probability distribution of the semantic-level behaviors of the first type of important nodes with the semantic-level behaviors in the decision tree of the ego vehicle two by two as a scene for closed-loop forward simulation, while in the simulation process of a planning period, the nodes in the risk tree are set as key nodes that need to update the confidence, and the confidence of other nodes is not updated in the planning period, and only the probability of the maximum confidence is considered as the initial confidence.

[0085] Step 5.4: Evaluate each scene for closed-loop forward simulation to select the best strategy. The total reward of each decision strategy is determined by the weighted sum of the rewards obtained by each scene selected in the previous step, and the reward function of each scene is determined by the comprehensive cost function F e including efficiency F s , safety F n and navigation F total , F total =-R:

[0086] F total =λ1F e +λ2F s +λ3F n

[0087] Wherein, λ1, λ2, λ3 are weight coefficients.

[0088]

[0089] Wherein, is a weight coefficient; N a is the number of semantic-level behaviors in the decision strategy; Δv p is the difference between the current speed v ego of the ego vehicle in the simulation and the preferred speed v pref , Δv p =|v ego -v pref |; Δv0 is the difference between the speed v lead of the detected front vehicle in the target lane and the speed v ego The difference, Δv, represents the traffic efficiency of the target lane. o =max(v ego -v lead ,0);Δv l Let Δv be the difference between the vehicle's preferred speed and the speed of the vehicle ahead in the target lane. l =|v lead -v pref |

[0090] Safety costs are assessed by detecting potential collisions and incorporating a Responsibility-Sensitive Safety (RSS) model. If the vehicle is currently in a hazardous state according to the RSS model, the cost is determined by adjusting the vehicle's speed v. ego Compared with the currently obtained safe speed range ( and The difference between the lower and upper bounds of the safe speed range is used to penalize the vehicle's current state.

[0091]

[0092] Among them, b c and b r Boolean values ​​used to determine whether a collision has occurred and whether the RSS model has been violated; N is the penalty coefficient when a collision occurs. s This represents the total number of simulated states for the vehicle's trajectory. This is an adjustable parameter.

[0093] Navigation cost function F n This is used to assess driver preferences and the consistency of decision-making strategies across different planning cycles.

[0094]

[0095] in, λ consist b is the correlation coefficient; navi b consist These are the matching degree between the driving strategy provided to the user and the navigation target, and the matching degree between the current decision strategy and the historical decision strategy.

[0096] In the overall cost function F total In the middle, the efficiency cost function F e Safety cost function F s It is generated based on each semantic-level behavior. A discount factor γ is introduced to adjust the future R, and the scenarios simulated in each closed-loop forward simulation during the planning cycle are evaluated. The semantic-level behavior strategy with the largest R is selected as the final decision strategy.

[0097] The beneficial effects of this invention are:

[0098] 1、The application proposes an intelligent driving cognitive method based on complex network theory, models a complex real driving scene, models other traffic participants as dynamic network nodes, realizes state cognition and risk correlation cognition of the nodes, and judges important nodes and extracts a risk tree. By abstracting a complex real driving environment into a dynamic network model, the behavior decision of an intelligent driving car is guided.

[0099] 2、The application combines complex network cognitive theory and POMDP to complete the behavior decision of an intelligent driving car. Compared with the behavior decision method using a finite state machine and RL, the proposed method can effectively operate in a highly interactive dynamic uncertain driving environment while ensuring the interpretability of the algorithm, and provides a new solution for the development of a high-level intelligent driving system. BRIEF DESCRIPTION OF DRAWINGS

[0100] Figure 1 A method flowchart of the application

[0101] Figure 2 A dynamic point model used by the cognitive theory based on a complex network

[0102] Figure 3 A moving point model used by the cognitive theory based on a complex network

[0103] Figure 4 A construction process of a self-car decision tree

[0104] Figure 5 A scene of closed-loop forward simulation simulation divided by a cognitive model DETAILED DESCRIPTION

[0105] The application will be further described below with reference to the drawings.

[0106] The application provides an intelligent driving behavior decision method combining a cognitive theory based on a complex network and a POMDP theory, which can realize safe, efficient and flexible behavior decision in a highly interactive dynamic uncertain driving environment. As shown in the figure, the method specifically includes the following steps: Figure 1

[0107] First, the driving environment is modeled by using the cognitive theory based on a complex network, the real driving environment is abstracted into a dynamic network model, and important nodes are determined and a risk tree is extracted.

[0108] (1) Based on the complex network theory, the self-car and other traffic participants are regarded as network nodes, and a dynamic complex network model is constructed:

[0109] G t =(P,E,W,Θ)​ t

[0110] where G t is a dynamic complex network model; P = {p1, p2,..., p N} is a set of nodes in the network, p i is a node in the network, N is the number of nodes in the network; E = {e1, e2,..., e K} is a set of edges of nodes in the network, e i is an edge of nodes in the network, K is the number of edges of nodes in the network; W = {w1, w2,..., w k} is a set of weights of edges; Θ is a movable area of nodes.

[0111] Θ is modeled as a smooth bounded surface:

[0112]

[0113] where Ω is the boundary of the smooth bounded surface.

[0114] The state of a node can be modeled as:

[0115]

[0116] where X i ∈ R m is the state vector of node p i ; U i ∈ R q is the input vector; A i is a dynamic matrix; B i is an input matrix of control quantity, then the output vector of node p i is:

[0117] Y i = f i (X i )

[0118] The weight function between node p i and node p j can be expressed as:

[0119] w ij = F(Y i , Y j )

[0120] 1) Based on the two-degree-of-freedom model of the vehicle and the moving point model, the dynamics of the node is modeled, and the related nodes are regarded as moving particles with a center of gravity, and a moving point model relative to the expected path is established to predict the future motion trajectory of the node in the absence of vehicle-to-vehicle communication and real-time state estimation means. Through the dynamic and kinematic modeling, the state cognition of each node is realized.

[0121] 2) Based on field theory, a variable Gaussian safety field model based on risk center transfer is proposed, revealing the dynamic field characteristics of nodes. Under static conditions, the safety field model can be described by a two-dimensional Gaussian function, thus the static safety field S... sta It can be represented as:

[0122]

[0123] Where x0 and y0 are the coordinates of the risk center (i.e., the origin of the coordinate system centered on the vehicle); C a a is the electric field strength coefficient; x and b y These are the radii of the major and minor axes of the ellipse (the ellipse is an enlarged version of the inscribed ellipse of the vehicle's frame, which can be equivalently represented by the vehicle's aspect ratio).

[0124] As the vehicle moves, the risk center will shift accordingly, and the new risk center will be O′(x′0, y′0):

[0125]

[0126] in, k is the velocity vector. v It is a regulating factor, and 0 < k v <1 (forward) or -1 < k v <0 (backward); β is Angle with the x-axis. Use a′ x and b′ y Let S represent the radii of the major and minor axes of the ellipse after the risk center shifts. Then the dynamic safety field S... dyn It can be represented as:

[0127]

[0128] (2) Based on the human driver’s sensitivity to distance and reaction time, the area around the node is divided into the first cognitive domain, the second cognitive domain and the outer space.

[0129] The scope of the first cognitive domain is:

[0130] maxa′ x ≤s th1

[0131] s th1 =t c1 ·v e

[0132] Among them, s th1 The first threshold; t c1 The first cognitive reaction time of a human driver; v e This represents the maximum approach speed of other nodes in the environment.

[0133] The scope of the second cognitive domain is:

[0134] s th1 <maxa′ x ≤s th2

[0135] s th2 =t c2 ·v e

[0136] Among them, s th2 The second threshold; t c2 This refers to the second cognitive reaction time of a human driver. The space outside the second cognitive domain is the extra-domain space.

[0137] Within the framework of a variable Gaussian safety field, a risk perception function is established between nodes:

[0138]

[0139] in, It is node p i At node p j Field strength at the location, It is node p j scalar velocity, direction angle θ i，j It is node p j velocity vector With node p i field intensity vector The included angle (clockwise is positive), k c This is the risk perception adjustment coefficient.

[0140] (3) Based on the established model, generate a complex network and evaluate important nodes and extract risk trees.

[0141] 1) Set the vehicle node as the master node, connect the master node and other nodes in the first cognitive domain, calculate the weight of the corresponding connection, and sort them according to the weight. For ease of calculation, if there are multiple nodes in the connection direction with the master node, take the nearest node as the node in the first cognitive domain, and regard other nodes that are far away as nodes in the second cognitive domain. Do the same for the nodes in the second cognitive domain.

[0142] 2) In the first cognitive domain, select environmental nodes in sequence according to the weight sequence determined in the previous step, calculate their weights with other nodes and sort them, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0143] 3) Select the environmental nodes in the second cognitive domain and the nodes in the first cognitive domain, and connect the node pairs with the largest weights.

[0144] 4) In the second cognitive domain, select environmental nodes in sequence, calculate their weights relative to other nodes, sort them, and connect the node pairs with the largest weights (if the weight is lower than the set threshold, do not connect). If the edge already exists, do not connect it again.

[0145] 5) Evaluate key nodes and generate a risk tree.

[0146] Let node p i The set of all neighboring nodes is N. i ; Node strength r i For node p i The sum of the weights of its neighboring nodes is then:

[0147]

[0148] The average strength of all nodes in the network is defined as the network strength. <r>Then we have:

[0149]

[0150] The node strength r i The ratio of the sum of the intensities of all nodes is defined as To assess the importance of a node, we have:

[0151]

[0152] Will The largest node is defined as an important node. Important nodes in the first cognitive domain are defined as first-class important nodes (nodes that are not the main node), and important nodes in the second cognitive domain are defined as second-class important nodes. The tree formed by the main node, first-class important nodes, and second-class important nodes is defined as a risk tree.

[0153] The above steps enable the construction of a cognitive model based on complex networks, while simultaneously evaluating important nodes in the network and generating a risk tree.

[0154] Furthermore, applying the POMDP theory to... Figure 1 Modeling the behavioral decision-making process in [the context].

[0155] Let ε t The driving environment at time t, centered on the vehicle, includes the cognitive model in the cognitive module of the diagram, traffic signals, and the grid map of static obstacles. Let z represent the state of vehicle i∈V at time t. When i=0, it represents the vehicle itself. If the state variable lacks a superscript, it represents all vehicles; if it lacks a subscript, it represents all time points. The observation value received by the vehicle itself is denoted as z. t And use the observations to evaluate the true state quantity x t The input to the behavioral planning layer is denoted as... <z t , ε t The output is a set of continuous discrete state parameters, denoted as... Where H represents the time length of a planning period (in this invention, the planning period is set to 5 seconds), x t+i This represents the state of all vehicles at time t+i.

[0156] The POMDP model is defined as a set of six elements.<X,A,Z,T,O,R> Where X, A, and Z represent the state space, behavior space, and observation space, respectively; T represents the behavior a taken by the vehicle at time t. t After ∈A, the state probability transition function T(x) at consecutive times t-1 a t x t )=p(x t |x t-1 a t ), O is the probability function between state and observation O(x t , z t ) = p(z t | x t ), which reflects the uncertainty of the driving environment. R(x t-1 , a t ) is the reward function, which represents the real-time reward obtained by the ego vehicle after taking action a t-1 ∈ A in state x t ∈ X.

[0157] When the ego vehicle is in a real driving scene, the state space can be represented as where N is the number of traffic participants, t is the current time, and the state of the vehicle includes pose information, the strength of the nodes in the cognitive model constructed based on complex network theory, and hidden states of the vehicle (such as behavior intention) that cannot be directly observed.

[0158] Then T can be represented as The ego vehicle is distinguished from other traffic participants, and it is assumed that the transformation of its state is instantaneous, so T can be further transformed into:

[0159]

[0160] where, is the assumed driver model, which reflects the upper-level decision-making process of other traffic participants; is the state probability transition model of other traffic participants, which reflects the kinematics process of other traffic participants executing decisions. To simplify the solution, the present application sets it as a deterministic kinematics model.

[0161] In order to eliminate uncertainty, the POMDP model makes decisions on the belief space , which is the probability distribution of state X. The belief of the belief space will be updated from the initial belief b t after the traffic participant takes action a t and receives observation z t-1 . The belief update is performed using the Bayes formula:

[0162]

[0163] where η is the normalization factor. The observation of other vehicles is obtained by the ego vehicle's perception and prediction module, so:

[0164]

[0165] Assuming that the observation process is independent of each other, the belief update process can be further written as:

[0166]

[0167] From the above equations, it can be seen that although the state transition process of each traffic participant is independent of each other, the belief update process is an interactive process, because and use the state quantities and observations of all traffic participants, so that the belief update process is an interactive process, and the hidden states of other traffic participants are updated step by step in the belief update process.

[0168] Therefore, in the POMDP model, the purpose of decision-making is to find a best decision strategy π in the belief space * , so that the ego vehicle can obtain the maximum reward expectation after taking a series of discrete behaviors in the planning period, that is:

[0169]

[0170] where t0 is the starting time of the current planning; γ is the discount factor of the reward R, 0≤γ≤1; is the initial belief; b t is the belief at time t. Starting from the starting time and the initial belief, the action space A and the observation space O are expanded step by step in the planning period t H , and the nodes of the belief tree are generated step by step in the belief space. By using the Bellman equation for each node, the optimal strategy is found:

[0171]

[0172] where is the optimal utility function of the belief; Q * (b, a) is the optimal value that can be obtained after taking a certain behavior in the belief space. Then the behavior and observation value corresponding to the optimal strategy of each node can be found by searching the belief tree:

[0173]

[0174]

[0175] Taking the above steps can obtain the state quantities corresponding to the optimal strategy of each node, including the belief the behavior and the observation value that is For let the final decision strategy

[0176] Finally, according to the established model, the behavior decision of the ego vehicle is made.

[0177] For the convenience of problem solving, the application abstracts high-level semantic behaviors (such as straight, left turn, right turn, etc.) from the driver model of the traffic participant Replace the complex behavior action Therefore, in the decision-making process of the whole model, only the semantic behavior input needs to be controlled to complete the behavior decision of the ego vehicle.

[0178] (1) The behavior sequence of the ego vehicle is defined as a set of ordered decision trees in time sequence, and the nodes of the tree are pre-defined semantic behaviors generated in time steps within the planning period. The decision goal of the ego vehicle is to find a set of semantic behaviors from the root node within the whole planning period, including a series of nodes, taking the leaf node of the previous stage as the root node of the next stage, and repeating the generation process of the decision tree. The decision-making space Φ of the whole decision can be defined by a limited number of semantic strategies.

[0179] (2) The trajectory of the other traffic participant observed by the perception module is denoted as the observation value z t , assuming that the observation value is noiseless and the state of the ego vehicle is known. Through the observation model , the hidden intention of the other traffic participant can be further evaluated, and the hidden intention of the other traffic participant can be mapped to its corresponding semantic behavior, i.e. is deterministic, the random behavior of the other traffic participant can be further converted into a confidence state, and the multiple trajectories of the other traffic participant can be predicted through the prediction module, denoted as the initial confidence of the confidence space The confidence of the confidence space is updated through the confidence update formula established by the POMDP model, thereby completing the construction of the observation space Z and the confidence space .

[0180] (3) The nodes of the ego vehicle decision tree and the semantic behaviors of the other traffic participants are defined as a scenario for closed-loop forward simulation, and the probability of each scenario can be obtained through the joint distribution of the confidence probability of each traffic participant.

[0181] (4) Each scenario that is simulated by closed-loop forward simulation is evaluated, and the semantic behavior strategy with the largest R is selected as the final decision strategy The total reward of each decision strategy is determined by the weighted sum of the rewards obtained by each scenario selected in the previous step, and the reward function of each scenario is determined by the comprehensive cost function F e containing efficiency F s , safety F n and navigation F total :

[0182] F total = λ1F e + λ2F s + λ3F n

[0183] where λ1, λ2, λ3 are weight coefficients; F total = -R.

[0184]

[0185] where, are weight coefficients; N a is the number of semantic level behaviors in the decision strategy; Δv p is the difference between the current speed v ego of the ego vehicle and the preferred speed v pref in the decision strategy, Δv p = |v ego - v pref |; Δv o is the difference between the speed v lead of the detected front vehicle in the target lane and the speed v ego of the ego vehicle, which represents the traffic efficiency of the target lane, Δv o = max(v ego - v lead , 0); Δv l is the difference between the preferred speed of the ego vehicle and the speed of the front vehicle in the target lane, Δv l = |v lead - v pref |.

[0186] The safety cost is evaluated by detecting whether a collision occurs and introducing an RSS (responsibility-sensitive safety) model, and if the ego vehicle is currently in a dangerous state in the RSS, the current state of the ego vehicle is penalized by taking the difference between the speed v ego of the ego vehicle and the currently obtained safety speed interval (L and U are the lower bound and the upper bound of the safety speed interval, respectively):

[0187]

[0188] where b c and b r are Boolean values for determining whether a collision occurs and the RSS model is violated; is the penalty coefficient when a collision occurs; N s is the total number of ego vehicle trajectory simulation states; is an adjustable parameter.

[0189] The navigation cost function F n to evaluate the consistency of the driver's preference and different planning period decision strategies:

[0190]

[0191] wherein, λ consist is the correlation coefficient; b navi , b consist respectively are the matching degree of the driving strategy provided by the user and the navigation target and the matching degree between the current decision strategy and the historical decision strategy.

[0192] In the total cost function F total , the efficiency cost function F e , the safety cost function F s are generated according to each semantic level behavior. At the same time, the discount coefficient γ=0.75 is introduced to adjust the future R, and the optimal ego behavior decision strategy is selected by evaluating each closed loop forward-backward simulation scenario in the planning period as the input of the trajectory planning and control module.

[0193] As Figure 2 shown, the dynamics model used by the cognitive theory based on complex network proposed by the application is a two-degree-of-freedom vehicle model, which can be expressed as:

[0194]

[0195] wherein, the state variables of the model are the yaw rate r and the lateral velocity U y ; δ is the front wheel steering angle; F yf and F yr are respectively the front lateral force and the rear lateral force generated in driving; a and b are respectively the front axle center of gravity and the rear axle center of gravity; m is the vehicle mass; I z is the yaw moment of inertia; the longitudinal velocity U x is considered as a time-varying parameter. When the tire cornering characteristics are in the linear range, the model can be expressed as:

[0196]

[0197] wherein, C f and C r are respectively the front wheel cornering stiffness and the rear wheel cornering stiffness.

[0198] As Figure 3 shown, the kinematics model used by the cognitive theory based on complex network proposed by the application is a moving point model, which can be expressed as a moving point model by regarding the relevant nodes as moving particles with the center of gravity, and establishing the moving point model relative to the expected path:

[0199]

[0200] where, Δψ and e are the heading angle deviation and lateral path deviation, respectively; s is the distance along the desired path; and κ is the curvature of the desired path.

[0201] As shown in Figure 4 , the behavior sequence of the ego vehicle is defined as a set of decision trees generated in time sequence, the nodes of the tree are the predefined semantic level behaviors generated in time steps within a planning period, including K (straight), R (right turn), and L (left turn), and the decision goal of the ego vehicle is to find a set of semantic level behaviors starting from the root node within the whole planning period, take the leaf node of the last planning period as the root node of the next planning period, and thus perform the generation process of the decision tree in a cycle, so as to define the whole decision behavior space Φ with a limited number of semantic level strategies.

[0202] The size of the decision space will increase exponentially with the depth of the decision tree, in order to reduce the computational complexity, the decision tree is pruned in the present application. In real scenarios, the driver generally does not frequently change the semantic level behavior within the same planning period, therefore, it is set that each period contains at most one operation of changing the semantic behavior. At the same time, the cognitive model based on complex network is used to further prune the decision tree, when the risk weight between the ego vehicle and the direction of a node in the first cognitive domain is greater than the set threshold, the lane changing operation in the corresponding direction is prohibited (as shown in the dashed box in Figure 4 ), and by changing the size of the threshold, different driving styles of the ego vehicle can be reflected. The duration of each semantic level behavior is set to 1 s, and the depth of the tree is set to 5, so that a planning period is 5 s, and the decision tree is reconstructed and updated at a frequency of 0.5 s.

[0203] As shown in Figure 5 , the probability of each scenario can be obtained through the joint probability distribution of the confidence estimates of each traffic participant, therefore, the number of scenarios for simulation will increase exponentially with the number of traffic participants, which will make the complexity of the algorithm difficult to calculate, but in actual driving scenarios, we often do not need to pay attention to all vehicles. In order to solve this problem, the cognitive theory established in the present application is used to prune the number of scenarios for simulation, the evaluation selection range is set to the nodes within the second cognitive domain (maintain straight when there is no other traffic participant in the second cognitive domain), and the probability distribution of the semantic level behavior of the first type of important node is combined with the semantic level behavior of each node in the decision tree of the ego vehicle to form a scenario for closed loop forward simulation. At the same time, in the simulation process of a planning period, the nodes in the risk tree are set as the key nodes whose confidence needs to be updated, the confidence of other nodes is not updated within the planning period, and only the probability of the maximum confidence is considered as the initial confidence. The update frequency of the simulation is set to 0.2 s.

[0204] Based on the above method, the application further provides a vehicle electronic device, which is a control device or a storage device.

[0205] In summary, the application is an intelligent driving behavior decision method that fuses complex network theory and partially observable Markov decision process. By constructing a cognitive model, the complex driving scene is abstracted into a dynamic network model. Since POMDP is used, compared with the mainstream behavior decision method using finite state machine and RL, the proposed method can be more suitable for real driving environment with dynamic interaction while ensuring interpretability, providing a new solution for the development of high-level intelligent driving system.

[0206] The above series of detailed descriptions are only specific descriptions of the feasible embodiments of the application, and are not used to limit the protection scope of the application. Any equivalent means or changes without departing from the technology of the application shall be included in the protection scope of the application.< / r> < / r>

Claims

1. An intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process, characterized in that, Comprising the following steps: Step 1: dynamically modeling the driving environment based on the cognitive theory of complex networks; Step 2: dividing the nodes into the first cognitive domain, the second cognitive domain and the out-of-domain space according to the sensitivity of human drivers to distance and reaction time; Step 3: generating a complex network and evaluating important nodes and extracting a risk tree according to the model obtained in step 1; Step 4: modeling the decision-making process of the ego vehicle by applying the POMD theory to eliminate the uncertainty in the driving environment; The step 4 specifically comprises: Step 4.1: Set ε t represents the driving environment centered on the ego vehicle at time t, including the cognitive model modeled in the previous step, the traffic signal, the grid map of static obstacles, etc. represents the state of vehicle i∈V at time t, i = 0 represents the ego vehicle, and when the state quantity lacks a superscript, it represents all vehicles, and when it lacks a subscript, it represents all time points; the observation value received by the ego vehicle is denoted as z t , and the observation value is used to evaluate the true state quantity x t ; the input of behavior decision is denoted as <z t ,ε t >, and the output is a set of continuous and discrete state parameters, denoted as where H represents the length of a planning period. Step 4.2: POMDP model is defined as a set of six elements <X, A, Z, T, O, R>, where X, A and Z are state space, action space and observation space, respectively; T is the state probability transition function T(x t-1 ,a t ,x t ) = p(x t |x t-1 ,a t ), O is the probability function between state quantity and observation quantity O(x t ,z t ) = p(z t |x t ), by representing state quantity and observation value in the form of probability distribution, to reflect the uncertainty in the driving environment, R(x t-1 ,a t ) is the reward function, which represents the real-time reward obtained by the ego vehicle after taking action a t ∈ A in state x t-1 ∈ X; When the ego vehicle is in a real driving scenario, the state space can be represented as where N is the number of traffic participants, t is the current time, the state quantity of the vehicle includes pose information, node strength constructed in the complex network model, and hidden state of the vehicle that cannot be directly observed; T can be expressed as T can be further transformed to be wherein, is a hypothetical driver model reflecting the upper-level decision-making process of other road users; is a state probability transition model of other road users reflecting the kinematic process of other road users executing decisions, which is set as a motion point model for the purpose of simplifying the solution of the model; To eliminate the uncertainty, the POMDP model makes decisions in belief space which is a probability distribution over states X, the belief of the belief space will be updated after the traffic actor takes an action a t and receives an observation z t The initial belief b t-1 is updated using the Bayes formula for belief update: Wherein, η is a normalization factor, the observation value of other vehicles can be obtained by the perception and prediction module of the ego vehicle, then: Assuming that the observation process is independent of each other, the confidence update process is further written as: As can be seen from the above formula, although the state transition process of each traffic participant is independent of each other, and The state quantity and observation value of all traffic participants are used in the confidence update process, so that the confidence update process is an interactive process, thereby gradually updating the hidden state of other traffic participants in the confidence update process. In the POMDP model, the goal of decision is to find an optimal decision policy π in the belief space B * such that the maximum expected reward is achieved by the ego vehicle after a series of discrete actions are taken in a planning period, then we have: where t0is the current planned start time; γ is a discount factor for the reward R, taken as 0≤ γ≤ 1; The initial belief is initialized, and from the start time and the initial belief, the action space A and the observation space O are expanded step by step, and the nodes of the belief tree are generated step by step in the belief space. By using the Bellman equation for each node, the optimal policy can be found: H The initial belief is initialized, and from the start time and the initial belief, the action space A and the observation space O are expanded step by step, and the nodes of the belief tree are generated step by step in the belief space. By using the Bellman equation for each node, the optimal policy can be found: wherein, is the optimal utility function for the confidence; Q * (b, a) is the optimal value that can be obtained after taking a certain action in the confidence space, then the optimal strategy corresponding to each node can be found by the confidence tree corresponding to the action and observation value: Taking the above steps can get the state quantity corresponding to the optimal strategy of each node, including confidence, the behavior taken and the observation value, that is For Let The final decision strategy can be obtained Step 5: making behavior decisions of the ego vehicle according to the model built in step 4.

2. The intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process according to claim 1, characterized in that, The step 1 specifically comprises establishing a dynamic complex network model, specifically: Based on the theory of complex networks, the ego vehicle and other traffic participants are regarded as network nodes to construct a dynamic complex network model: G t = (P, E, W, Θ) t Wherein, G t is a dynamic complex network model; P={p1, p2,..., p N} is a set of nodes in the network, p i is a node in the network, N is the number of nodes in the network; E={e1, e2,..., e K} is a set of edges of nodes in the network, e i is an edge of a node in the network, K is the number of edges of nodes in the network; W={w1, w2,..., w k} is the weight of the edge; Θ is the movable area of the node; Θ is modeled as a smooth bounded surface: Wherein, Ω is the boundary of the smooth bounded surface.

3. The intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process according to claim 1, characterized in that, The step 1 further comprises establishing a dynamic model, specifically: Based on the two-degree-of-freedom vehicle model and the motion point model, the nodes are dynamically modeled: where the state variables of the model are the yaw rate r and the lateral velocity U y ; δ is the front wheel steering angle; F yf and F yr are the front and rear lateral forces generated during driving; the longitudinal velocity U x is considered as a time-varying parameter, and when the tire cornering characteristics are in the linear range, the model can be expressed as: The relevant nodes are regarded as motion particles with a center of gravity to establish a motion point model relative to the expected path: Wherein, Δψ and e are the heading angle deviation and lateral path deviation respectively; s is the distance along the expected path; κ is the curvature of the expected path.

4. The intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process according to claim 1, characterized in that, The step 1 further comprises establishing a variable Gaussian safety field model, specifically: A variable Gaussian safety field model based on risk center transfer is proposed according to the field theory, and the static safety field is described by a two-dimensional Gaussian function: where C a is the field strength coefficient; a x and b y are the radii of the major and minor axes of an ellipse, respectively, which is a magnification of the inscribed ellipse of the vehicle frame, and is equivalent to the aspect ratio of the vehicle. When the vehicle moves, the risk center will shift, and the new risk center is O'(x'0, y'0): where, is the velocity vector; k v is a tuning factor, and has 0 < k v < 1, indicating forward, or -1 < k v < 0 indicates backward; β is the angle with the x-axis, then the dynamic safety field can be expressed as:

5. The intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process according to claim 1, characterized in that, In the step 2: The range of the first cognitive domain is: max a' x ≤s th1 s th1 = t c1 · v e where s th1 is a first threshold value; t c1 is a first cognitive reaction time of a human driver; v e is a maximum approach speed of other nodes in the environment; The range of the second cognitive domain is: s th1 <maxa' x ≤s th2 s th2 = t c2 • v e wherein s th2 is a second threshold value; t c2 is a second cognitive reaction time of a human driver; The space outside the second cognitive domain is defined as the out-of-domain space; In the framework of the variable Gaussian safety field, the risk cognition function between nodes is established: where, is the field strength at node p i is the scalar velocity at node p j is the field strength at node p is the scalar velocity at node p j is the field strength at node p i,j is the scalar velocity at node p j is the field strength at node p is the field strength at node p i is the field strength at node p is the field strength at node p c is the risk perception adjustment factor.

6. The intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process according to claim 1, characterized in that, The step 3 specifically comprises: Step 3.1: evolution of the complex network: 1) Set the ego vehicle node as the main node, connect the main node and other nodes in its first cognitive domain, calculate the weight of the corresponding edges, and sort them according to the weight, if there are multiple nodes in the connection direction with the main node, take the nearest node as the node in the first cognitive domain, and other nodes with longer distance are regarded as nodes in the second cognitive domain, and the nodes in the second cognitive domain are processed in the same way; 2) In the first cognitive domain, select the environment nodes in sequence according to the weight sequence determined in the previous step, calculate the weight of the nodes with other nodes and sort them, connect the node pair with the largest weight, if the weight is lower than the set threshold, do not connect, if the node pair with the largest weight already exists, do not repeat the connection; 3) Select the environment nodes in the second cognitive domain and the nodes in the first cognitive domain, and connect the node pair with the largest weight; 4) In the second cognitive domain, the environmental nodes are selected in turn, the weights of which are calculated relative to other nodes and sorted, the node pair with the largest weight is connected, if the weight is lower than the set threshold, it is not connected, if the node pair with the largest weight already exists, it is not repeated; Step 3.2: Evaluate the important nodes and generate a risk tree, including: 1) Set of all adjacent nodes to node p i N i ; Node strength r i is the sum of the weights of its adjacent nodes, then: i ​ 2) The average of all the nodes strength in the network is defined as the network strength <r>Then there is:< / r> 3) The node strength r i The ratio of the strength of a node to the sum of the strengths of all nodes is defined as To assess the importance of a node, one has: Will The largest node is defined as an important node; important nodes in the first cognitive domain are defined as first-class important nodes; and important nodes in the second cognitive domain are defined as second-class important nodes. 4) The tree generated by the main node, the first type of important node and the second type of important node is defined as a risk tree.

7. The intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process according to claim 1, characterized in that, The step 5 specifically includes: Step 5.1: Abstract high-level semantic-level behaviors from the model of traffic participants High-level semantic-level behaviors include going straight, turning left, and turning right, replacing complex behavior actions Therefore, in the decision-making process of the entire model, only the semantic-level behavior input by the ego vehicle needs to be controlled to complete the behavior decision-making of the ego vehicle; The behavior sequence of the ego vehicle is defined as a set of ordered decision trees generated in time sequence, the nodes of the tree are predefined semantic level behaviors generated in time steps within the planning period, and the decision goal of the ego vehicle is to find a set of semantic level behaviors starting from the root node within the entire planning period, taking the leaf node of the previous stage as the root node of the next stage, and so on, so that the entire decision behavior space Φ can be defined by a limited number of semantic level strategies; The size of the decision space will increase exponentially with the depth of the decision tree, and the decision tree is pruned; in a real scene, the driver generally does not frequently change the semantic level behavior within the same planning period, so it is set that each period contains at most one operation of changing the semantic behavior, and the cognitive model based on the complex network established in step 1 is used for further pruning of the decision tree, when the risk weight between the ego vehicle and the direction of a node in the first cognitive domain is greater than the set threshold, the lane changing operation in the corresponding direction is prohibited, and the size of the threshold can reflect different driving styles of the ego vehicle; Step 5.2: Record the trajectories of other traffic participants as observations z t , assuming that the observations are noise-free, the ego state is known, and the observation model The partially hidden intentions of other traffic participants can be further evaluated, while the hidden intentions of other traffic participants can be mapped to their corresponding semantic-level behaviors, i.e. is deterministic, the behavior of other traffic participants can be further converted into a confidence state, predicting multiple trajectories of other traffic participants, which is recorded as the initial confidence of the confidence space The confidence space is updated by the confidence update formula established in step 4, thus completing the construction of the observation space Z and the confidence space Step 5.3: Define the semantic level behaviors of the ego vehicle and other traffic participants as a scene for closed-loop forward simulation, and the probability of each scene is obtained through the joint distribution of the confidence probability of each traffic participant; The constructed cognitive model is used to prune the number of simulation scenes, the evaluation selection range is set to the nodes in the second cognitive domain, the ego vehicle keeps straight when there is no other traffic participant in the second cognitive domain, the probability distribution of the semantic level behavior of the first type of important node is combined with the semantic level behavior in the decision tree of the ego vehicle to form a scene for closed-loop forward simulation, and in the simulation process of a planning period, the nodes in the risk tree are set as key nodes whose confidence needs to be updated, the confidence of other nodes is not updated within the planning period, and only the probability of the maximum confidence is considered as the initial confidence. Step 5.4: Evaluate each scenario that is simulated in the forward simulation for closed loop to pick the best strategy; the total reward for each decision strategy is determined by the weighted sum of the rewards obtained from each scenario selected in the previous step, the reward function for each scenario is determined by the combined cost function F e that includes efficiency F s , safety F n , and navigation F total , F total = -R: F total = λ1F e + λ2F s + λ3F n Wherein, λ1, λ2, λ3 are weight coefficients.

8. The intelligent driving behavior decision-making method fusing complex network theory and partially observable Markov decision process according to claim 7, characterized in that, The steps 5.4, efficiency F e , security F s and navigation F n are designed as follows, respectively: wherein, is a weight coefficient; N a is the number of semantic level behaviors in the decision strategy; Δv p is the difference between the current speed of the ego vehicle and the preferred speed in the simulation, Δv p = |v ego -v pref |; Δv o is the difference between the detected preceding vehicle and the ego vehicle in the target lane, indicating the passing efficiency of the target lane, Δv o = max(v ego -v lead , 0); Δv l is the difference between the preferred speed of the ego vehicle and the preceding vehicle speed in the target lane, Δv l = |v lead -v pref |; The safety cost is evaluated by detecting whether a collision occurs and introducing the RSS model, which represents the responsibility-sensitive safety. If the ego vehicle is currently in a dangerous state in the RSS model, the ego vehicle speed v ego The current state of the ego vehicle is penalized by subtracting the safety speed interval currently obtained: where b c and b r are Boolean values that determine whether a collision and a violation of the RSS model, respectively, have occurred; is a penalty coefficient in case of collision; N s is the total number of states of the ego trajectory simulation; is a tunable parameter; Navigation cost function F n to assess the consistency of the driver's preferences and different planning period decision strategies: wherein, λ consist is a correlation coefficient; b navi , b consist are the matching degree of the driving strategy provided by the user and the navigation target and the matching degree between the current decision strategy and the historical decision strategy, respectively.

9. A control apparatus for a vehicle characterized by comprising: The control device is provided with the decision method of any one of claims 1-8.

10. A storage device for a vehicle, characterized by comprising: The storage device is provided with the program of the decision method of any one of claims 1-8.

Images