CN115257746B - A lane-changing decision control method for autonomous driving vehicles considering uncertainty - Google Patents

Abstract

Translated fromChinese

本发明涉及一种考虑不确定性的自动驾驶汽车换道决策控制方法，包括：构建状态空间及动作空间，建立状态转移方程；建立自车、他车的观测空间模型及信念空间模型、他车预测轨迹的不确定性模型；设定奖励函数，结合POMDP模型，求解自车的决策状态点集、并解耦为横向空间决策集及纵向时间决策集；分别确定横向、纵向可行驶边界，引入道路边界约束，车速约束及障碍物距离约束，划分出横向可行驶区域及参考路径、纵向可行驶区域及参考速度曲线；根据决策结果，车辆规划模块输出相应车辆最优轨迹，使车辆按照最优轨迹行驶。与现有技术相比，本发明充分考虑他车预测轨迹的不确定性，决策出的自车状态点集更加稳定可靠，能有效增强车辆驾驶的舒适性及安全性。

The present invention relates to a lane-changing decision control method for an autonomous driving vehicle considering uncertainty, including: constructing a state space and an action space, establishing a state transfer equation; establishing an observation space model and a belief space model of the self-vehicle and other vehicles, and an uncertainty model of the predicted trajectory of the other vehicle; setting a reward function, combining a POMDP model, solving the decision state point set of the self-vehicle, and decoupling it into a lateral space decision set and a longitudinal time decision set; respectively determining the lateral and longitudinal drivable boundaries, introducing road boundary constraints, vehicle speed constraints and obstacle distance constraints, dividing the lateral drivable area and reference path, longitudinal drivable area and reference speed curve; according to the decision result, the vehicle planning module outputs the optimal trajectory of the corresponding vehicle, so that the vehicle drives according to the optimal trajectory. Compared with the prior art, the present invention fully considers the uncertainty of the predicted trajectory of other vehicles, and the state point set of the self-vehicle determined by the decision is more stable and reliable, which can effectively enhance the comfort and safety of vehicle driving.

Description

Uncertainty-considered automatic driving automobile lane change decision control method

Technical Field

The invention relates to the technical field of automatic driving automobile control, in particular to an automatic driving automobile lane change decision control method considering uncertainty.

Background

The decision planning is one of the key parts of automatic driving, firstly, multi-sensor information is fused, then task decision is carried out according to driving requirements, then, on the premise that possible obstacles can be avoided, a plurality of selectable safety paths between two points are planned through specific constraint conditions, and an optimal path is selected from the paths to serve as a vehicle driving track.

At present, a decision module of a vehicle integrally gathers required information, and makes reasonable driving decisions through comprehensive analysis, and a planning module of the vehicle outputs safe and comfortable vehicle tracks according to decision results. In existing research, representative decision methods include finite state machines, behavioral trees, markov processes, etc. However, most decision methods consider less uncertainty problems caused by dynamic changes of environment and vehicle states in a future period, and most decision methods simply output behavior instructions, so that the rationality of decision results cannot be guaranteed, even the decision results are discontinuous in time, and therefore, a problem of difficult solution is brought to a track planning module, and poor safety and comfort of vehicles and poor user experience are easily caused.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide the automatic driving automobile lane change decision control method taking uncertainty into consideration, and the reliability of a decision result is improved by taking uncertainty of a predicted track of the other automobile into consideration, so that the comfort and safety of driving the automobile are enhanced.

The aim of the invention can be achieved by the following technical scheme: an automatic driving automobile lane change decision control method considering uncertainty comprises the following steps:

s1, based on a perceived vehicle state and a predicted other vehicle state, a state space and an action space are built by combining a vehicle dynamics model, and a state transition equation is built;

S2, respectively establishing an observation space model and a belief space model of the own vehicle, an observation space model and a belief space model of the other vehicle and an uncertainty model of a predicted track of the other vehicle;

s3, setting a reward function, combining the steps S1 and S2 to construct and obtain a POMDP (PARTIALLY OBSERVABLE MARKOV DECISION PROCESS, part of observable Markov decision process) model, and solving a decision state point set of the vehicle;

S4, decoupling the decision state point set of the own vehicle into a transverse space decision set and a longitudinal time decision set;

S5, determining a transverse drivable boundary, introducing road boundary constraint, vehicle speed constraint and obstacle distance constraint, and dividing a transverse drivable region and a reference path;

Determining a longitudinal drivable boundary, introducing road boundary constraint, vehicle speed constraint and barrier distance constraint, and dividing a longitudinal drivable region and a reference speed curve;

s6, outputting an optimal track of the corresponding vehicle by the vehicle planning module according to the transverse drivable area and the reference path, the longitudinal drivable area and the reference speed curve which are divided in the step S5, so that the vehicle can run according to the optimal track.

Further, the step S1 specifically includes the following steps:

S11, simplifying the motion of a vehicle into the motion of particles in a Frenet coordinate system by using a simplified vehicle kinematic model so as to construct a state space and an action space of the vehicle;

S12, using the state space and the action space to establish state transfer functions of the own vehicle and the other vehicle.

Further, the step S11 specifically includes the following steps:

s111, acquiring the perceived vehicle position, longitudinal speed, longitudinal acceleration and lateral speed information of the self-vehicle to form a state space of the self-vehicle;

the method comprises the steps of collecting predicted course angle, vehicle position, longitudinal speed, longitudinal acceleration and lateral speed information of the other vehicle, and combining the length and width of the vehicle body to form a state space of the other vehicle;

s112, setting a discrete transverse acceleration sequence and a discrete longitudinal acceleration sequence as an action space.

Further, the specific process of step S12 is as follows: the evolution process of the states of the own vehicle and the other vehicle is assumed to be independent, a state transition equation of the own vehicle is obtained according to a simplified vehicle motion model, and the state transition equation of the other vehicle is expressed by probabilities of different states at the next moment.

Further, the observation function of the observation space model in the step S2 is divided into a vehicle and another vehicle, wherein the vehicle observation function is represented by a boolean value, if the next state exists, the vehicle observation function is set to 1, otherwise, the vehicle observation function is set to 0;

The observation function of the other vehicle accords with Gaussian distribution;

And in the step S2, an uncertainty model of the prediction track of the other vehicle is constructed by adopting a multi-element Gaussian distribution.

Further, the step S3 specifically includes the following steps:

S31, setting a reward function comprising a safety reward function R_safe, a comfort reward function R_comfort and an efficiency reward function R_efficiency, wherein the safety reward function R_safe comprises a collision reward function R_colli and a distance reward function R_dis;

The comfort rewards function R_comfort includes a speed dependent comfort rewards function R_speed, a lateral speed action penalty function R_vlat, and a continuity indicator R_continuity;

The efficiency rewards function R_efficiency includes a target task rewards function R_lane and a target speed function R_{v_tar};

s32, solving a POMDP model by adopting a deterministic sparse observable tree (DETERMINED SPARSE PARTIALLY Observable Tree, DESPOT) method based on the constructed state space, action space, belief space and rewarding function to obtain a decision state point set of the own vehicle, wherein the content of the decision state point set comprises position information of the own vehicle and speed and acceleration information at different moments.

Further, the step S31 is specifically to design the collision reward function R_colli according to the probability of collision, that is, giving a penalty when the collision probability exceeds the set threshold;

step S31 is to design a distance rewarding function R_dis according to a collision time (TTC) model, and set upper and lower limits of a maximum safe distance;

In the step S31, the influencing factors of the speed-related comfort bonus function R_speed include the current longitudinal speed vlon_cur, the longitudinal acceleration alon_cur, the lateral speed vlat_cur, and the longitudinal speed vlon_next, the longitudinal acceleration alon_next, and the lateral speed vlat_next at the next moment;

In the step S31, the lateral velocity action penalty function R_vlat is related to the lateral velocity only;

In the step S31, the continuity indicator R_continuity is positively correlated with the position change amounts of the two previous and subsequent decisions, and the larger the position change amounts of the two previous and subsequent decisions, the larger R_continuity is.

Further, the lateral space decision set in step S4 includes location information of the own vehicle; the longitudinal time decision set includes speed and acceleration information of the own vehicle.

Further, in the step S5, when the lateral drivable boundary is selected, the distance from the node of the lateral drivable boundary to the obstacle is greater than the safety distance, and the lateral drivable boundary value does not exceed the original boundary of the structured road, and the distance of a safety threshold is maintained;

The establishment process of the transverse reference path specifically comprises the following steps:

Firstly, constructing a transverse optimization problem, designing a transverse cost function to evaluate each node, and taking a point at the minimum position of the transverse cost function as an optimal solution, wherein the constraint of the transverse optimization problem is that a course angle is between the minimum and maximum wheel corners, the objective function of the transverse optimization problem is the transverse cost function, and the transverse cost function C_node-h is a weighted sum of a first distance cost C_d, a first safety cost C_o and a continuity cost C_c;

Calculating the distance from the node to the target state point as a first distance cost; calculating a distance from the node to the obstacle as a first security cost; calculating the position change rate between the front node and the rear node as a continuity cost; combining the corresponding weight coefficients, and calculating to obtain a transverse cost function;

And finally, taking the value at the minimum of the transverse cost function, and connecting the lines to form a transverse expected reference path.

Further, the principle of determining the longitudinal drivable boundary in step S5 is as follows: the longitudinal drivable boundary coincides with the parking space coordinates of the obstacle;

the constraints of the longitudinal runnability boundary are: the s-t curve of the longitudinal drivable boundary does not exceed the s-t curve representing the highest and lowest average vehicle speed;

the establishment process of the longitudinal reference speed curve specifically comprises the following steps:

Firstly, constructing a longitudinal optimization problem, designing a longitudinal cost function to evaluate each node, and taking a point at the minimum position of the longitudinal cost function as an optimal solution, wherein the longitudinal cost function C_node-z is a weighted sum of a second distance cost C_d-ref, a second safety cost C_o-ref and a speed change cost C_v;

Calculating the distance from the node to the state point connecting line to be used as a second distance cost; calculating the square of the difference between the distance from the node to the obstacle on the s-axis and the distance threshold value from the obstacle as a second safety cost; calculating the change rate of the reference speed as the speed change cost; calculating to obtain a longitudinal cost function by combining the corresponding weight coefficients;

And finally, taking the value at the minimum of the longitudinal cost function, and forming a longitudinal expected reference speed curve by connecting the two points.

Compared with the prior art, the method fully considers the uncertainty of the predicted track of the own vehicle, establishes the uncertainty model of the predicted track of the own vehicle by establishing the observation space model and the belief space model of the own vehicle and the other vehicle, and obtains the decision state point set of the own vehicle based on the POMDP model by combining the set reward function. Therefore, the determined vehicle state point set is more stable and reliable, the comfort and the safety of driving the vehicle are enhanced, and the vehicle driving method is well applicable to unmanned driving under typical driving conditions such as vehicle following, lane changing and autonomous overtaking.

After the decision state point set of the own vehicle is obtained by solving, the decision state point set is decoupled into a transverse space decision set and a longitudinal time decision set, and the reliability of a decision result can be further ensured by respectively dividing a transverse drivable region and a reference path and dividing a longitudinal drivable region and a reference speed curve through introducing road boundary constraint, vehicle speed constraint and barrier distance constraint.

Drawings

FIG. 1 is a schematic flow chart of the method of the present invention;

FIG. 2 is a schematic diagram of a lane change decision process according to an embodiment;

FIG. 3 is a schematic diagram of uncertainty of a predicted trajectory of a host vehicle in an embodiment;

FIG. 4 is a schematic diagram of a lateral search determining a lateral travelable boundary in an embodiment;

FIG. 5 is a diagram of a lateral travelable region and a lateral desired reference path in an embodiment;

Fig. 6 is a graph of a longitudinal travelable region and a longitudinal desired reference speed profile in an embodiment.

Detailed Description

The invention will now be described in detail with reference to the drawings and specific examples.

Examples

As shown in fig. 1, an uncertainty-considered automatic driving automobile lane change decision control method comprises the following steps:

s1, based on a perceived vehicle state and a predicted other vehicle state, a state space and an action space are built by combining a vehicle dynamics model, and a state transition equation is built;

S2, respectively establishing an observation space model and a belief space model of the own vehicle, an observation space model and a belief space model of the other vehicle and an uncertainty model of a predicted track of the other vehicle;

s3, setting a reward function, combining the steps S1 and S2 to construct a POMDP model, and solving a decision state point set of the vehicle;

S4, decoupling the decision state point set of the own vehicle into a transverse space decision set and a longitudinal time decision set;

S5, determining a transverse drivable boundary, introducing road boundary constraint, vehicle speed constraint and obstacle distance constraint, and dividing a transverse drivable region and a reference path;

Determining a longitudinal drivable boundary, introducing road boundary constraint, vehicle speed constraint and barrier distance constraint, and dividing a longitudinal drivable region and a reference speed curve;

s6, outputting an optimal track of the corresponding vehicle by the vehicle planning module according to the transverse drivable area and the reference path, the longitudinal drivable area and the reference speed curve which are divided in the step S5, so that the vehicle can run according to the optimal track.

The channel change decision process of the embodiment is shown in fig. 2, and includes:

1. Based on the perceived state of the own vehicle and the predicted state of the other vehicle, constructing a state space and an action space according to the simplified vehicle dynamics model, and establishing a state transition equation;

the method specifically comprises the following steps:

11 Using simplified vehicle kinematic model to simplify the motion of vehicle into particle motion in Frenet coordinate system, to construct state space and action space of vehicle;

111 State space expression is:

S＝[state_ego,state₁,state₂,...,state_n]

Wherein state_ego is a state space of the vehicle, time is a time stamp, (s_ego,l_ego) is a position of the vehicle, vlon_ego is a longitudinal speed of the vehicle, acc_ego is a longitudinal acceleration of the vehicle, and vlat_ego is lateral speed information of the vehicle;

state_n is a state space of the other vehicle n, (length, width) is a length and width of the vehicle body of the other vehicle n, θ is a predicted course angle of the other vehicle, (s, l) is a position of the other vehicle, and v_n is speed information of the other vehicle;

112 A discrete lateral acceleration sequence and a discrete longitudinal acceleration sequence are set as the action space:

A＝[Acc_lon,Vel_lat]

Vel_vlat＝{-2.0,-1.5,-1.0,-0.5,0.0,0.5,1.0,1.5,2.0}

Acc_lon＝{-3.0,-2.5,-2.0,-1.5,-1.0,-0.5,0.0,0.5,1.0,1.5,2.0}

Wherein Vel_vlat is the lateral acceleration sequence; acc_lon is the longitudinal acceleration sequence;

12 Using the state space and the action space to establish the state transfer functions of the own vehicle and other vehicles;

the specific establishment process is as follows:

121 Assuming that the evolution processes of the states of the own vehicle and the other vehicle are mutually independent, obtaining a state transition equation of the own vehicle according to a simplified vehicle motion model, wherein the state transition equation of the other vehicle is expressed by probabilities of different states at the next moment;

the calculation formula of the state transition equation of the bicycle is as follows:

the calculation formula of the state transfer equation of the other vehicle is as follows:

2. establishing an observation space model and a belief space model of the own vehicle and the other vehicle, and establishing an uncertainty model of a prediction track of the other vehicle, wherein the observation space comprises information which can be observed by an intelligent vehicle environment perception positioning system, and comprises position, speed and course angle information of the own vehicle and position, speed and course angle information of the other vehicle;

The method comprises the following specific steps:

21 The state of the self-vehicle is completely considerable, and the observation function of the self-vehicle comprises coordinates and course angles, and accords with Gaussian distribution;

an observation function of an observation space is established, the self-vehicle observation function is represented by a Boolean value, if the next state exists and is set to be 1, otherwise, the self-vehicle observation function is set to be 0:

the observation function of the vehicle is:

the observation function of the other vehicle accords with Gaussian distribution, and the calculation formula is as follows:

where μ_s,l,θ represents the observed value, σ_s,l,θ represents the variance;

The process of calculating the position uncertainty through the observation model comprises the following steps:

22 Constructing an uncertainty model of the vehicle pose of the other vehicle by adopting a multi-element Gaussian score;

The uncertainty calculation formula of the self-vehicle pose of the other vehicle is as follows:

wherein Φ is a state transition matrix, M represents noise in a state transition process, Z is covariance of the noise, Σ is a covariance matrix of the system, X is a state quantity, and the expression mode is as follows:

the Gaussian distribution covariance of each state quantity on the prediction track point of the other vehicle is obtained through uncertainty modeling and is shown in figure 3;

3. setting a reward function, and solving a decision state point set of the vehicle based on the POMDP model;

31 The content included in the set bonus function: safety index R_safe, comfort index R_comfort, and efficiency index R_efficiency;

32 A bonus function for designing a security indicator): considering the collision reward function R_colli and the distance reward function R_dis, the specific calculation formula is:

R_safe＝R_colli+R_dis

33 Designing a collision reward function R_colli according to the probability of collision, giving a penalty when the collision probability exceeds a threshold, and specifically solving the collision reward function as follows:

Where w_colli = 100 is a weight value, P_safe is a safety threshold;

designing a reward function R_dis of the distance between the front vehicles according to a collision occurrence time (TTC) model, setting upper and lower limits of the maximum safe distance, and specifically calculating the reward function of the distance between the front vehicles as follows:

wherein d_max、d_min is the maximum safety distance and the minimum safety distance between the own vehicle and other vehicles, and w_dis =20 is the weight value of the distance rewards;

34 Design comfort rewards function: considering the speed-related comfort reward function R_speed, the lateral speed action penalty function R_vlat and the continuity indicator R_continuity, the comfort reward function is written:

R_comfort＝R_speed+R_vlat+R_continuity

the solving formula of the speed-related comfort rewarding function R_speed is as follows:

The current longitudinal speed is vlon_cur, the current longitudinal acceleration is alon_cur, the current lateral speed is vlat_cur, the longitudinal speed at the next moment is vlon_next, the longitudinal acceleration at the next moment is alon_next, the lateral speed at the next moment is vlat_next,w_speed =15, and the weight coefficient of the speed is the same;

The lateral velocity action penalty function is calculated as:

R_vlat＝-vlat²×w_vlat

wherein w_vlat = 12 is the weight coefficient of the lateral velocity;

the continuity index calculation formula of the result is:

Wherein t represents the time of each step of the decision result; m and n represent the total step numbers of the front decision result and the rear decision result respectively, (s_{last_loop,t},l_{last_loop,t}) and (s_{cur_loop,t},l_{cur_loop,t}) are the position information of the last decision result and the current decision result respectively, and w_con =12 is the weight coefficient of the continuity index;

35 Design efficiency bonus function R_efficiency: considering the target task rewards function R_lane and the target speed function R_{v_tar}, the efficiency rewards function is written as:

R_efficiency＝R_lane+R_{v_tar}

Wherein the target task rewarding function is R_lane＝-|l_lane-l_ego|×w_lane;

The target speed function is R_{v_tar}＝-|v_tar-v_ego|×w_{v_tar};

w_{v_tar} = 12 is a weight coefficient for the target speed;

36 Solving a decision action sequence, dividing the decision into a longitudinal space decision set of a transverse space decision set, and specifically solving a POMDP model by adopting a deterministic sparse observational tree method to obtain a vehicle decision state point set comprising vehicle position information and speed and acceleration information at different moments;

4. Decoupling the decision state point set into a transverse space decision set and a longitudinal time decision set, wherein the transverse space decision set comprises the position information of the own vehicle at each moment; the longitudinal time decision set comprises speed and acceleration information of each moment of the vehicle;

5. determining a transverse drivable boundary, introducing road boundary constraint, vehicle speed constraint and barrier distance constraint, and dividing a transverse drivable region and a reference path;

when the lateral drivable boundary is selected, the distance from the node of the boundary to the obstacle is larger than the safety distance, and the specific expression is as follows:

Wherein d_obs is the distance from the current node to the obstacle, and when the distance is smaller than a certain safety distance, the node is regarded as unsafe, l_k is the lateral coordinate of the node, and l_min≤l_k≤l_max should be satisfied;

at the same time, it should also be possible to meet the boundary values l_min and l_max, at which the driving boundary value does not exceed the original boundary of the structured road, and to maintain a distance of the safety threshold d_safe:

the calculated transverse travelable area is shown in fig. 4;

establishing a transverse optimization model in a transverse drivable area to obtain a transverse reference path, wherein the transverse reference path comprises the following specific steps of:

Constructing a transverse optimization problem, designing a transverse cost function to evaluate each node, wherein the point at the minimum of the transverse cost function is the optimal point;

The constraint is heading angle constraint:

wherein, θ'_min、θ'_max is the minimum value and maximum value of the change of the course angle of the vehicle movement respectively;

the transverse cost function is used as an objective function, and the expression is:

C_node-h＝w_dC_d+w_oC_o+w_cC_c

Wherein, C_d is the first distance cost of the node, C_o is the first security cost, C_c is the continuity cost, and the corresponding weight functions are w_d＝0.35,w_o＝0.45,w_c =0.2, respectively;

the first distance cost function C_d is expressed as:

Wherein (s_state,l_state) is the location of the target state point and (s_k,l_k) is the location of the extension node;

The first security cost function C_o is expressed as:

Where d_obs represents the distance of the node from the nearest obstacle and d_max represents the distance threshold from the obstacle;

the cost function C_c representing continuity is expressed as:

Wherein, l_i,l_i+1,l_i+2 is the lateral position of the front and rear three nodes;

Finally, a value connecting line at the minimum position of the transverse cost function is taken to form a transverse expected reference path;

The transverse desired reference path formed in this embodiment is represented in fig. 5 by the hollow dots;

6. Determining a longitudinal drivable boundary, introducing road boundary constraint, vehicle speed constraint and barrier distance constraint, and dividing a longitudinal drivable region and a reference speed curve;

The method for determining the longitudinal drivable boundary specifically comprises the following steps:

Ensuring that the boundary s_border in the s-axis direction satisfies the formula:

the s_border constraint is: s_{min_t}≤s_border≤s_{max_t}, wherein s_{max_t} and s_{min_t} are the upper and lower boundaries of the average speed, respectively;

the method for establishing the longitudinal reference speed curve specifically comprises the following steps:

constructing a longitudinal optimization problem, designing a longitudinal cost function to evaluate each node, wherein the point at the minimum of the longitudinal cost function is the optimal point; the longitudinal cost function is calculated as:

C_node-z＝w_d-refC_d-ref+w_o-refC_o-ref+w_vC_v

Wherein C_d-ref is the second distance cost from the node to the state point connection, C_o-ref is the second security cost of the node, C_v is the speed change cost, and the corresponding weight coefficients are w_d-ref＝0.4,w_o-ref＝0.4,w_v =0.2, respectively.

The second distance cost function is:

c_d-ref＝(s_i-s_ref)²

Wherein s_ref is the position of the node on the two-state point connection line;

the second safety cost function is:

Where s_obs represents the distance along the s-axis of the node from the nearest obstacle and s_max represents the distance threshold from the obstacle;

The cost function of the speed change is:

and finally, connecting values at the minimum of the longitudinal cost function to form a longitudinal expected reference speed curve.

The longitudinal reference velocity profile formed in this embodiment is shown in fig. 6.

In summary, the technical scheme considers the behavior interaction characteristics of the own vehicle and the traffic participants and the uncertainty of the predicted track of the other vehicle, the decision process is more stable, and the decided own vehicle state point set is more reasonable and reliable, so that the following vehicle planning module can accurately and rapidly solve, and the comfort and safety of vehicle driving are effectively enhanced.

Claims (6)

Translated fromChinese

1.一种考虑不确定性的自动驾驶汽车换道决策控制方法，其特征在于，包括以下步骤：1. A lane-changing decision control method for an autonomous driving vehicle considering uncertainty, characterized in that it comprises the following steps:S1、基于感知到的自车状态及预测的他车状态，结合车辆动力学模型，构建状态空间及动作空间，并建立状态转移方程；S1. Based on the perceived state of the vehicle and the predicted state of other vehicles, combined with the vehicle dynamics model, the state space and action space are constructed, and the state transfer equation is established;S2、分别建立自车的观测空间模型及信念空间模型、他车的观测空间模型及信念空间模型、他车预测轨迹的不确定性模型；S2, respectively establish the observation space model and belief space model of the own vehicle, the observation space model and belief space model of the other vehicle, and the uncertainty model of the predicted trajectory of the other vehicle;S3、设定奖励函数，结合步骤S1和S2构建得到POMDP模型，求解出自车的决策状态点集；S3, set the reward function, combine steps S1 and S2 to construct the POMDP model, and solve the decision state point set of the vehicle;S4、将自车的决策状态点集解耦为横向空间决策集及纵向时间决策集；S4, decoupling the decision state point set of the vehicle into a lateral spatial decision set and a longitudinal temporal decision set;S5、确定横向可行驶边界，引入道路边界约束，车速约束及障碍物距离约束，划分出横向可行驶区域及参考路径；S5. Determine the lateral drivable boundary, introduce road boundary constraints, vehicle speed constraints and obstacle distance constraints, and divide the lateral drivable area and reference path;确定纵向可行驶边界，引入道路边界约束，车速约束及障碍物距离约束，划分出纵向可行驶区域及参考速度曲线；Determine the longitudinal drivable boundary, introduce road boundary constraints, vehicle speed constraints and obstacle distance constraints, and divide the longitudinal drivable area and reference speed curve;S6、根据步骤S5划分出的横向可行驶区域及参考路径、纵向可行驶区域及参考速度曲线，车辆规划模块输出相应车辆最优轨迹，使车辆按照最优轨迹行驶；S6, according to the lateral drivable area and reference path, longitudinal drivable area and reference speed curve divided in step S5, the vehicle planning module outputs the optimal trajectory of the corresponding vehicle, so that the vehicle drives according to the optimal trajectory;所述步骤S1具体包括以下步骤：The step S1 specifically includes the following steps:S11、使用简化的车辆运动学模型，将车辆的运动简化为Frenet坐标系中质点的运动，以构建自车的状态空间及动作空间；S11. Use a simplified vehicle kinematic model to simplify the vehicle's motion into the motion of a mass point in the Frenet coordinate system to construct the state space and action space of the ego vehicle;S12、利用状态空间和动作空间，建立转移自车及他车的状态转移函数；S12, using the state space and action space, establish a state transfer function for transferring the self-vehicle and other vehicles;所述步骤S11具体包括以下步骤：The step S11 specifically includes the following steps:S111、采集感知的自车的车辆位置、纵向速度、纵向加速度、侧向速度信息，构成自车的状态空间；S111, collecting the perceived vehicle position, longitudinal speed, longitudinal acceleration, and lateral speed information of the ego vehicle to form a state space of the ego vehicle;采集预测的他车的航向角、车辆位置、纵向速度、纵向加速度、侧向速度信息，结合车身长宽，构成他车的状态空间；Collect the predicted heading angle, vehicle position, longitudinal speed, longitudinal acceleration, and lateral speed information of the other vehicle, and combine it with the length and width of the vehicle body to form the state space of the other vehicle;S112、设定离散的横向加速度序列及纵向加速度序列，以作为动作空间；S112, setting a discrete lateral acceleration sequence and a longitudinal acceleration sequence as an action space;所述步骤S2中观测空间模型的观测函数分为自车和他车，其中，自车观测函数由布尔值表示，若下一状态存在，则设置为1，否则设置为0；The observation function of the observation space model in step S2 is divided into the own vehicle and other vehicles, wherein the own vehicle observation function is represented by a Boolean value, which is set to 1 if the next state exists, otherwise it is set to 0;他车的观测函数符合高斯分布；The observation function of the other car conforms to the Gaussian distribution;所述步骤S2中他车预测轨迹的不确定性模型采用多元高斯分布构建；The uncertainty model of the predicted trajectory of the other vehicle in step S2 is constructed using multivariate Gaussian distribution;所述步骤S3具体包括以下步骤：The step S3 specifically comprises the following steps:S31、设定奖励函数包括安全性奖励函数R_safe、舒适性奖励函数R_comfort及效率性奖励函数R_efficiency，其中，安全性奖励函数R_safe包括碰撞奖励函数R_colli及距离奖励函数R_dis；S31, setting the reward function to include a safety reward function R_safe , a comfort reward function R_comfort , and an efficiency reward function R_efficiency , wherein the safety reward function R_safe includes a collision reward function R_colli and a distance reward function R_dis ;舒适性奖励函数R_comfort包括速度相关舒适性奖励函数R_speed、侧向速度动作惩罚函数R_vlat以及连续性指标R_continuity；The comfort reward function R_comfort includes the speed-related comfort reward function R_speed , the lateral speed action penalty function R_vlat and the continuity index R_continuity ;效率性奖励函数R_efficiency包括目标任务奖励函数R_lane和目标速度函数R_{v_tar}；The efficiency reward function R_efficiency includes the target task reward function R_lane and the target speed function R_{v_tar} ;S32、基于构建的状态空间、动作空间、信念空间及奖励函数，采用确定性稀疏可观测树的方法，求解POMDP模型，得到自车的决策状态点集，其中，决策状态点集的内容包括自车的位置信息以及不同时刻的速度和加速度信息。S32. Based on the constructed state space, action space, belief space and reward function, the deterministic sparse observable tree method is used to solve the POMDP model and obtain the decision state point set of the vehicle, where the content of the decision state point set includes the position information of the vehicle and the speed and acceleration information at different times.2.根据权利要求1所述的一种考虑不确定性的自动驾驶汽车换道决策控制方法，其特征在于，所述步骤S12的具体过程为：假设自车和他车状态的演进过程是相互独立的，根据简化的车辆运动模型，得到自车的状态转移方程，他车的状态转移方程则由下一时刻不同状态的概率表达。2. According to the uncertainty-taking lane-changing decision control method for an autonomous driving vehicle according to claim 1, the specific process of step S12 is as follows: assuming that the evolution processes of the states of the self-vehicle and the other vehicle are independent of each other, the state transfer equation of the self-vehicle is obtained according to a simplified vehicle motion model, and the state transfer equation of the other vehicle is expressed by the probabilities of different states at the next moment.3.根据权利要求1所述的一种考虑不确定性的自动驾驶汽车换道决策控制方法，其特征在于，所述步骤S31具体是根据碰撞发生的概率设计碰撞奖励函数R_colli，即当碰撞概率超过设定阈值时，给予惩罚；3. The lane changing decision control method for an autonomous driving vehicle considering uncertainty according to claim 1, characterized in that the step S31 specifically designs a collision reward function R_colli according to the probability of collision, that is, when the collision probability exceeds a set threshold, a penalty is given;所述步骤S31具体是根据碰撞发生时间(TTC)模型设计距离奖励函数R_dis，并设定最大安全距离的上下界限；The step S31 specifically designs a distance reward function R_dis according to a time to collision (TTC) model, and sets upper and lower limits of the maximum safety distance;所述步骤S31中，速度相关舒适性奖励函数R_speed的影响因素包括当前的纵向速度vlon_cur、纵向加速度alon_cur、侧向速度vlat_cur以及下一时刻的纵向速度vlon_next、纵向加速度alon_next、侧向速度vlat_next；In the step S31, the influencing factors of the speed-related comfort reward function R_speed include the current longitudinal speed vlon_cur , longitudinal acceleration alon_cur , lateral speed vlat_cur and the longitudinal speed vlon_next , longitudinal acceleration alon_next , lateral speed vlat_next at the next moment;所述步骤S31中，侧向速度动作惩罚函数R_vlat只与侧向速度有关；In the step S31, the lateral speed action penalty function R_vlat is only related to the lateral speed;所述步骤S31中，连续性指标R_continuity与前后两次决策的位置变化量正相关，前后两次决策的位置变化量越大、则R_continuity越大。In the step S31, the continuity index R_continuity is positively correlated with the position change between two decisions. The greater the position change between the two decisions, the greater the R_continuity .4.根据权利要求1所述的一种考虑不确定性的自动驾驶汽车换道决策控制方法，其特征在于，所述步骤S4中横向空间决策集包括自车的位置信息；纵向时间决策集包括自车的速度及加速度信息。4. According to the uncertainty-taking lane changing decision control method of an autonomous driving vehicle according to claim 1, it is characterized in that the lateral spatial decision set in step S4 includes the position information of the vehicle; the longitudinal time decision set includes the speed and acceleration information of the vehicle.5.根据权利要求1所述的一种考虑不确定性的自动驾驶汽车换道决策控制方法，其特征在于，所述步骤S5中横向可行驶边界在选择时，横向可行驶边界的节点到障碍物的距离大于安全距离，同时横向可行驶边界值不超过结构化道路的原始边界，并保持一个安全阈值的距离；5. The lane change decision control method for an autonomous driving vehicle considering uncertainty according to claim 1, characterized in that, when selecting the lateral drivable boundary in step S5, the distance from the node of the lateral drivable boundary to the obstacle is greater than the safety distance, and the lateral drivable boundary value does not exceed the original boundary of the structured road, and maintains a distance of a safety threshold;横向可行驶区域参考路径的建立过程具体为：The specific process of establishing the reference path of the lateral drivable area is as follows:首先构建横向优化问题，设计横向成本函数对每个节点进行评价，取横向成本函数最小处的点作为最优解，其中，所述横向优化问题的约束为航向角在最小及最大车轮转角之间，所述横向优化问题的目标函数为横向成本函数，所述横向成本函数为第一距离成本C_d，第一安全成本C_o及连续性成本C_c的加权和；First, a lateral optimization problem is constructed, and a lateral cost function is designed to evaluate each node, and the point where the lateral cost function is the minimum is taken as the optimal solution, wherein the constraint of the lateral optimization problem is that the heading angle is between the minimum and maximum wheel turning angles, and the objective function of the lateral optimization problem is the lateral cost function, which is the weighted sum of the first distance cost C_d , the first safety cost_Co and the continuity cost C_c ;通过计算从节点到目标状态点的距离，以作为第一距离成本；计算从节点到障碍物的距离，以作为第一安全成本；计算从前后节点之间的位置变化率，以作为连续性成本；结合对应的权重系数，计算得到横向成本函数；By calculating the distance from the node to the target state point as the first distance cost; calculating the distance from the node to the obstacle as the first safety cost; calculating the position change rate between the front and rear nodes as the continuity cost; combining the corresponding weight coefficients, the lateral cost function is calculated;最终取横向成本函数最小处的值，连线形成横向期望参考路径。Finally, the value at the minimum of the lateral cost function is taken, and the lines are connected to form the lateral expected reference path.6.根据权利要求5所述的一种考虑不确定性的自动驾驶汽车换道决策控制方法，其特征在于，所述步骤S5中纵向可行驶边界的确定原则为：纵向可行驶边界与障碍物车位置坐标重合；6. The lane-changing decision control method for an autonomous driving vehicle considering uncertainty according to claim 5, characterized in that the principle for determining the longitudinal drivable boundary in step S5 is that the longitudinal drivable boundary coincides with the position coordinates of the obstacle vehicle;纵向可行驶边界的约束条件为：纵向可行驶边界的s-t曲线不超过表示最高、最低的平均车速的s-t曲线；The constraint condition of the longitudinal drivable boundary is: the s-t curve of the longitudinal drivable boundary does not exceed the s-t curve representing the highest and lowest average vehicle speeds;纵向可行驶区域参考速度曲线的建立过程具体为：The specific process of establishing the reference speed curve of the longitudinal drivable area is as follows:首先构建纵向优化问题，设计纵向成本函数对每个节点进行评价，取纵向成本函数最小处的点作为最优解，其中，所述纵向成本函数为第二距离成本C_d-ref，第二安全成本C_o-ref及速度变化成本C_v的加权和；Firstly, a longitudinal optimization problem is constructed, a longitudinal cost function is designed to evaluate each node, and the point where the longitudinal cost function is the minimum is taken as the optimal solution, wherein the longitudinal cost function is the weighted sum of the second distance cost C_d-ref , the second safety cost_Co-ref and the speed change cost C_v ;通过计算从节点到状态点连线的距离，以作为第二距离成本；计算节点到障碍物在s轴上的距离与距障碍物的距离阈值之差的平方，以作为第二安全成本；计算参考速度的变化率，以作为速度变化成本；结合对应的权重系数，计算得到纵向成本函数；The distance from the node to the state point is calculated as the second distance cost; the square of the difference between the distance from the node to the obstacle on the s-axis and the distance threshold to the obstacle is calculated as the second safety cost; the rate of change of the reference speed is calculated as the speed change cost; combined with the corresponding weight coefficient, the longitudinal cost function is calculated;最终取纵向成本函数最小处的值，连线形成纵向期望参考速度曲线。Finally, the value at the minimum of the longitudinal cost function is taken, and the lines are connected to form the longitudinal expected reference speed curve.