Research On Behavior Decision-making And Motion Planning For Intelligent Vehicle In Urban Uncertainty Environments

The behavioral decision-making and motion planning module is the core part of the intelligent driving system.The urban traffic environment is highly dynamic and uncertain,especially the future movements of other traffic participants in the perception information.Ignoring these uncertainties will affect the safety of intelligent vehicles.However,ensuring the safety of all possible uncertainties would make intelligent vehicles very conservative.Therefore,considering multiple design requirements such as safety,comfort,efficiency and practicability,this paper designed a behavioral decision-making and motion planning algorithm that comprehensively considers the uncertainty of perception information,multiple driving requirements and vehicle dynamics.Firstly,a decision planning framework that takes into account the uncertainty of perceptual information is designed.Decoupling the behavioral decision layer and the motion planning layer,the behavioral decision layer is responsible for considering the uncertainty of future motions of other traffic participants,and building partially observable markov decision processes(POMDP)to reason about future scenarios at relatively coarse resolution to generate initial behavioral trajectories.This interface can effectively improve the consistency of the decision planning system compared to the traditional single decision command.The motion planning layer integrates multiple driving demand objectives and vehicle dynamics,decouples the motion planning problem under complex constraints into a multi-objective lateral optimization problem and a multi-objective longitudinal optimization problem,and solves a more refined and high-quality motion trajectory.Secondly,aiming at the behavioral decision-making problem in the urban uncertain environment,taking the most challenging lane changing scenarios and intersection scenarios as application scenarios,a behavioral decision-making model considering the uncertainty of perception information is designed.Simplified modeling of the driving behavior under these two operating scenarios is carried out,and the uncertainty and collision risk in the scenarios are analyzed.The uncertainty of perception information considered in this paper includes the uncertainty of the driving intention of other traffic vehicles and the uncertainty of the future predicted trajectory.Under the premise of considering the influence of these uncertainties,POMDP models are established for two working scenarios,including state space,action space,state transition model and reward function.Aiming at the problem of large system dimensions caused by exploring low-confidence regions in the belief space in POMDP,the multi-strategy closed-loop simulation method is used to improve the solution efficiency.Starting from the initial belief space,the state of the own vehicle and other vehicles are continuously updated according to different semantic actions,and calculate the corresponding reward function until the current decision time domain is exceeded.In the decision-making process of changing lanes,the uncertainty probability distribution of the other vehicle is constructed through the Gaussian process,and then the control amount is calculated by the lateral model predictive control and the longitudinal Intelligent Driver Model(IDM)to update the own vehicle.The state of the policy is evaluated and selected according to the designed reward function.In the decision-making exploration process of the intersection scenario,when the collision time interval is less than the safety threshold,the longitudinal IDM model is used to consider the safe distance with other vehicles,and the next longitudinal acceleration command is generated.Through continuous circulation,the optimal state sequence in the decision time domain can be obtained,and a Responsibility Sensitive Safety(RSS)check mechanism is introduced in the backend to improve the robustness of behavioral decision-making.In the VTD simulation platform,the other vehicle cut-in scenario,lane change overtaking scenario and unprotected left turn scenario at the intersection are designed to validate the behavioral decision algorithm and compare it with the traditional method.The results show that the proposed POMDP model enables the autonomous vehicle to have the ability to estimate and evaluate the driving decision results,and can select the driving strategy with the largest reward function.Compared with the traditional reactive decision-making method,it can take into account both safety and driving efficiency.Thirdly,for the lateral multi-objective optimization algorithm,the initial lateral state sequence of the upper-layer POMDP decision is used as the optimization goal,and on this basis,the lateral safe driving corridor is established by the method of sampling and search,which can quickly find the convex space of lateral optimization,to limit the solution space of the lateral optimization problem and improve the solution efficiency.The vehicle two-degree-of-freedom model is introduced into the kinematics model under the Frenet road coordinate system to establish the lateral motion model.Multiple objective constraints in lateral optimization problems are analyzed,including equation constraints considering lateral motion model,vehicle stability envelope constraints and vehicle non-integrity constraints,and soft constraints on steering speed are introduced to improve the robustness of lateral optimization solutions.Using quadratic programming to efficiently solve lateral convex optimization problems,it can quickly converge to a local optimal solution.Simulation experiments show that it can generate high-quality path curves while ensuring real-time,security and stability.Fourthly,for the longitudinal multi-objective optimization algorithm,the initial longitudinal state sequence of the upper-level POMDP decision is used as the optimization objective,and the form of intersection of the future trajectories of the own vehicle and other vehicles is analyzed,including point overlap and line overlap.The longitudinal safe driving corridor is established based on the result of upper-level decision-making,this method can quickly find the convex space of longitudinal optimization,limit the solution space of longitudinal optimization problem,and improve the solution efficiency.The vehicle longitudinal model constraints are established with the third-order kinematics model in the Frenet road coordinate system.At the same time,the tire longitudinal and lateral force friction ellipse is introduced as the vehicle stability constraint,and the soft constraint of the longitudinal jerk is considered to improve the robustness of the longitudinal optimization solution.The longitudinal convex optimization problem is efficiently solved by quadratic programming,which can quickly converge to the local optimal solution.Simulation experiments show that this method can generate high-quality speed curves while ensuring real-time performance,safety and comfort.Finally,based on the intelligent driving electric vehicle test platform,the effectiveness of the designed algorithm was verified in urban lane change scenarios,intersection scenarios and multi-obstacle obstacle avoidance scenarios.The results show that the designed behavioral decision and motion planning algorithms improve the safety,driving efficiency and real-time performance of intelligent vehicles performing autonomous driving in the dynamic and uncertain environment of urban areas,and effectively promote the improvement of intelligent level of intelligent vehicles.