Research On Behavioral Decision Making For Intelligent Vehicles In Dynamic Urban Environments

Intelligent vehicles need to cope with other vehicles in dynamic urban environments,understand their motion intentions and perform reasonable driving behaviors.The corresponding decision making abilities mostly determine the driving performances of intelligent vehicles.Depending on“Research on Self-driving Vehicle Key Technologies and Test Platform in Real Urban Traffic Scenarios”?91420203?supported by National Natural Science Foundation of China,this thesis focuses on autonomous driving decision making problem in dynamic urban environments.The typical scenarios considered in this research are lane driving and uncontrolled intersections.To solve the decision making problem,a two-step decision making model is proposed based on finite state machine?FSM?and partial observed Markov decision process?POMDP?.In this model,the complex decision making problem is decomposed into two subproblems,including lateral decision making problem based on FSM and longitudinal decision making problem based on POMDP.Based on FSM,driving behaviors of intelligent vehicles are divided into finite driving states.Besides,in order to get the relations between theoretical model and real world scenarios,a special state called“hesitating”state is added.The longitudinal decision making process in POMDP considers uncertain driving intentions of other vehicles,which generates the velocity policies.In this decision making framework,intelligent vehicles interact with other vehicles by adjusting speeds.Therefore,the driving behaviors of intelligent vehicles are stable and predictable.For the problem of solving this decision making model,the policy generation and policy selection mechanism are built as an efficient solver.The optimal policy is selected by generating possible candidate policies,predicting the future evolution of the whole situation and calculating the rewards based on multiple objective functions.In this process,there are three subproblems,including driving behavior recognition,scenario prediction and reward functions design.For driving behavior recognition,the driving behaviors are divided into lateral driving behavior?e.g.,lane changing and turning right in an intersection?and longitudinal driving behavior?e.g.,acceleration,deceleration?.Due to the demand of decision making module,we select easy observed parameters for intelligent vehicles as training inputs.Besides,Hidden Markov model?HMM?with Gaussian mixture model?GMM?and random forest model are used to train the final classification model.In addition,the performance is evaluated to validate the accuracy of such models.For scenario prediction problem,based on the lateral behavior prediction model,the longitudinal intention prediction model is built by GMM and HMM in lane driving and uncontrolled intersections.Given the candidate policies,the scenario prediction model is built with current lateral state of intelligent vehicles,aiming to predict the evolution of the scenario in a given period.In this prediction process,the comfort constraints are used to restrict other vehicle's driving maneuvers.For the problem of designing reward functions,considering vehicle properties,the reward functions are developed in six factors including safety,economy,comfort,time efficiency,traffic law and task completion,which is easily understood by humans.Besides,the weighted-sum model is used in multiobjective optimization process.Except for selecting weights by experts,the voting methods are also used to get the proper weights.Finally,we perform simulations and real world experiments.The simulation framework is built with PreScan software.Based on this simulation system,lane driving and intersection handling experiments are performed.The results show that intelligent vehicles perform cooperative driving behaviors with other vehicles successfully.In addition,the decision making model is implemented in“Ray”,a real world intelligent vehicle platform.Automated driving test in Beijing 3^rd Ring Road shows that the decision making model is feasible and effective.