Research On Behavior Decision-making Approaches For Autonomous Vehicle In Urban Uncertainty Environments

Autonomous vehicles have great application value in modern transportation systems and extreme dangerous scenarios,which can also be served as the synthetically experiment platforms for related novel technologies.In recent years,autonomous vehicles have attracted extensive attention around the world.As the "brain" of autonomous vehicle,the behavior decision-making system determines the safety and reasonability of autonomous driving.Improving the intelligent level of behavior decision-making system is always the hotspot and intractable task in autonomous driving domain.Comparing with freeway and other traffic environments,urban traffic environments are extreme complex and uncertain.Its complexity are mainly caused by the intricate of road network topology,various types of traffic participants and road elements,and their complex interactive relationships.The uncertainty are mainly refer to the uncertainty of perceptional information,the difficulty to predict other vehicles' movement and so on.For the complexity and uncertainty,autonomous vehicles' behavior decision-making system should provide high reliable and safe results in time.However,most of the current behavior decision-making methods can only adapt to simple and certain driving environments,which cannot generate reasonable results in time.Targeting on this issue,this dissertation proposed one type of behavior decision-making method based ontology and markov theory.Three key technologies are studied in depth,i.e.driving scenario modeling,other vehicles' motion prediction and driving action generation.The specific research contents are summarized as follows:(1)Aiming to the non-accuracy decisions caused by existing methods'insufficiency representation ability to represent the multi-source heterogeneous information,a novel ontology-based driving scenario modeling method is proposed.First,the hierarchy and interactions among the information are studied carefully.A structural ontology model is proposed with considering the semantic relationships between the scenario elements,which modeled the driving scenario efficiently.Then,based on the semantical description of the driving scenario,a driving knowledge base and an online reasoning system are constructed.The prior driving knowledge are applied efficiently by this system.This method improved the traditional representation methods'(like the grid decomposition method)shortages,like poor understandable,non-comprehensive representation of elements' inter-relationships.The proposed method also overcame the difficulty to update the priori driving knowledge incrementally,which is caused by the highly coupled prior knowledge with reasoning phase.The proposed methods provide solid foundation for other vehicles' motion prediction and driving action generation.(2)Aiming to the unreasonable driving actions generated by the previous behavior decision-making methods,which are caused by the lack of other vehicles' motion prediction,this dissertation proposed one type of others' vehicles motion prediction method based on driving intention.Firstly,one type of other vehicles' driving intention prediction method is proposed based on hidden markov model and knowledge-based reasoning,which can predict other vehicles' driving intention adaptively under the dynamic traffic environments.Then,using the driving intention as the guidance information,a three-order Bezier curve based fast trajectory prediction method is proposed.At last,with the fully consideration of the uncertainty of drivers' driving behavior,other vehicles' motion state model is proposed based on the predicted trajectories.Based on the experimental results,the proposed prediction methods extended the prediction time horizon of driving intention prediction and the real-time capability of trajectory prediction.The precision and scenario adaptive ability for motion prediction are also improved.This part provide the pre-condition for the driving action generation under the dynamic traffic environments.(3)This dissertation proposed one type of driving action generation method,which combined Partially Observable Markov Decision Process(POMDP)and logical reasoning together.This method is targeting the unreasonable behavior decisions caused by the neglect of uncertainty and poor knowledge-based reasoning ability,which are the main drawbacks of the existing methods.fn order to consider the uncertainty of the perceptual information,a POMDP-based probabilistic model is proposed for driving action generation.Then,targeting on the poor knowledge-based reasoning ability and 'curse of dimensionality' issues of the POMDP model,two attention allocation characteristics of human drivers' visual behavior is analyzed,i.e.the goal-driven feature and the un-uniformity of the particle-size distribution feature.Based on these two features,a logical reasoning system with consideration of driving regulations and driving leading information.By the proposed system,the driving scenario can be reasoned efficiently and the state space's dimensionality can be compressed effectively.At last,the driving action is generation based on the creatively cooperation of the probabilistic model and the logical reasoning system.The results of our case study confirm that the proposed driving action generation method is superior to the traditional finite state machine based behavior decision-making methods on safety and efficiency.Based on the "Intelligent Pioneer II" autonomous vehicle,the effectiveness and rationality of the proposed behavior decision-making and scenario understanding approaches are verified in the real urban road environment.The following conclusions can be confirmed by the experimental results:the study improved the safety and efficiency of autonomous vehicles' decision-making ability under the urban complex and uncertainty environments;the intelligent level of the autonomous vehicles can also be improved efficiently.