Research On Human-Like Decision-Making And Planning Strategy For Intelligent Vehicles Based On Cognition Of Traffic Situation

Intelligent vehicles represent the trend and frontier of global automotive technology and have become the strategic commanding heights of the automotive industry.As the core component of intelligent vehicle technology,decision-making and planning system plays an important role,but also faces severe challenges: On the one hand,the interaction relationship between the elements of human-vehicle-environment coupling system is complex and has strong dynamic uncertainty.On the other hand,the driving tasks undertaken by the decisionmaking and planning system and the driving objectives it needs to consider are more diversified,while also needing to avoid excessive human-machine decision-making and planning differences.Learning from the wisdom of human drivers to correctly understand the situational evolution of the human-vehicle-environment system,and then establishing human-like decision-making and planning strategy through learning drivers’ driving behaviors,provides an effective way to address the above challenges.However,the transient spatial risk assessment methods commonly used at present are difficult to accurately describe the spatiotemporal evolution of traffic situations,and standardized objective risk prediction methods also fail to consider the different subjective risk cognitive characteristics of heterogeneous drivers.Furthermore,driving behaviors of human drivers exhibit strong randomness,discreteness,and time variation,making them extremely difficult to learn.Existing research lacks a systematic analysis and characterization of the internal driving behaviors’ generation mechanism,leading to difficulties in achieving human-like multiobjective collaborative optimization in decision-making and planning strategies.Additionally,the transfer and generalization capabilities of strategies are limited in stochastic and uncertain scenarios.To address the above key issues,this paper,which is relies on the project of National Natural Science Foundation of China Project “Research on Human-Like Decision-Making and Planning Strategy for Intelligent Vehicles under Mixed Traffic Environment Based on Dynamic Generation Mechanism of Driving Behavior(52172386)” and “Research on Human-machine Parallel Control Conflict Mechanism and Key Technology of Cooperative Co-piloting for Intelligent Vehicles(51775235)”,researches and establishes a human-like decision-making and planning strategy for intelligent vehicles based on cognition of traffic situation,aiming to improve the safety,comfort,traffic efficiency,social recognition,and user acceptance,trust,fitness of intelligent vehicles,and achieve high-quality collaboration between human wisdom and machine intelligence,so as to provide theoretical support and technical references for the technological development and industrial application of intelligent vehicles.This paper establishes a trajectory prediction model for traffic participants,and based on the trajectory prediction results and the driver’s risk cognition mechanism,predicts driving risks that conform to the driver’s cognitive characteristics,thereby achieving humanlike cognition of traffic situation with subjective driving risk prediction as the core.Based on the cognition result of traffic situation,the human-like decision-making and lane-changing trajectory planning strategy and the human-like car-following motion planning strategy are established,which achieve human-like decision-making for lane-changing and lane-keeping.Subsequently,human-like lane-changing trajectory planning and human-like car-following motion planning are implemented based on the decision-making results.The main research contents of this paper are as follows:Firstly,a traffic participant trajectory prediction model based on the hierarchical spatiotemporal encoding and parallel multimodal decoding is established.The hierarchical spatiotemporal encoder gradually encodes and aggregates complex scenario information from three levels: local region spatial interaction information,local region temporal dependence relationship,and global spatiotemporal information.The parallel multimodal decoder processes the local and global spatiotemporal information through two parallel decoders,thus decoding the multimodal predicted trajectory of predicted agents and the probabilities of being in each modal.Verification and testing results on the Argoverse dataset demonstrate that the established prediction model can accurately and efficiently characterize the interactive coupling relationships and dynamic evolution trends among the multiple elements in the human-vehicle-environment system.It exhibits good performance in singletrajectory prediction and comprehensive quality of multi-modal predicted trajectories.Secondly,a subjective driving risk prediction model based on the spatiotemporal distribution of the driver’s cognitive risk is established.The risk cognition mechanism of the driver is analyzed under the human-vehicle-environment coupling system,and the driver’s cognitive risk is regarded as the combination of the environmental objective risk and driver’s subjective cognition,then the framework of the subjective driving risk prediction is constructed.Based on the traffic participant trajectory prediction,the objective anisotropic risk field describing the objective environmental risk and the driver’s subjective spatiotemporal cognition field describing the driver’s risk cognitive characteristics are established,and then the two fields are combined to construct the spatiotemporal distribution features of the driver’s cognitive risk.These features are input into the subjective driving risk prediction model composed of convolutional neural network,attention mechanism and bidirectional long short-term memory network,and the subjective driving risk matching the cognitive characteristics of drivers is obtained.Participants are recruited to watch the driving videos collected from the real world,and the subjective driving risk levels reported by the participants are used as the labels of learning samples.The verification results prove that the proposed model is effective and advanced in the subjective driving risk prediction,and can realize the human-like cognition of traffic situation on intelligent vehicles.Thirdly,based on the results of human-like traffic situation cognition,a human-like decision-making and lane-changing trajectory planning strategy based on lane-changing spatiotemporal targets is constructed.The strategy includes two parts: human-like lanechanging spatiotemporal target decision and human-like lane-changing trajectory planning.The lane-changing spatiotemporal target decision part can not only generate the human-like binary decision results for lane-changing and lane-keeping behaviors but also provides multidimensional decision targets such as the lane-changing starting time,the lane-changing duration,the lane-changing target position,and the lane-changing target velocity,and then guides the follow-up human-like lane-change trajectory planning.The generation mechanism of the driver’s lane-changing behaviors is characterized as the decision and guidance mechanism of lane-changing spatiotemporal targets considering multi-objective collaborative optimization,and then the strategy framework is established through the semantic expression,deconstruction and reorganization of the lane-changing behavior generation mechanism,and the lane-changing spatiotemporal targets are defined.Through the generation of expected trajectory space,design of human-like reward functions,Boltzmann noise rational model,and maximum entropy inverse reinforcement learning,the human-like lane-changing spatiotemporal target decision is achieved.Under the guidance of the spatiotemporal targets,the human-like lane-changing trajectory planning part of the proposed strategy will output the complete lane-changing trajectory through the hierarchical spatiotemporal encoding and the sliding time window decoding.The verification results on the high D dataset indicate that the proposed strategy can output the lane-changing spatiotemporal targets that are highly consistent with those of human drivers,and achieve human-like lane-changing trajectory planning with adaptive lane-changing duration.Then,when the target lane given by the human-like lane-changing spatiotemporal target decision-making part is the current lane,the intelligent vehicle needs to perform lanekeeping and,in the presence of a preceding vehicle,needs to achieve car-following driving.The generation mechanism of driver’s car-following behaviors is characterized as a multiobjective collaborative optimization car-following mechanism considering the motion of the preceding vehicle,and then a human-like car-following motion planning strategy based on reinforcement learning and supervised learning hybrid driven is constructed.The multiobjective optimization in the car-following motion planning is realized by reinforcement learning with time sequence of actions.The driver car-following reference model based on Gaussian mixture regression and continuous hidden Markov model is used to guide the strategy to learn the driver’s following behavior with appropriate tendencies.The strategy can consider the dynamic changes of the future traffic environment through the preceding vehicle motion prediction.The verification results on the high D dataset prove the advancement of the strategy on human-like car-following motion planning.The strategy can properly reproduce the driver’s car-following behavioral characteristics,and also has advantages over the driver in car-following safety,tracking,comfort and moving efficiency.Finally,the integrated verification of the proposed strategy is conducted based on field driving data of human drivers.The field data collection platform that can accurately obtain information of the ego vehicle and traffic environment is established,and drivers are recruited to carry out field driving data collection tests on conventional urban roads and urban expressways,and the naturalistic driving dataset Driver-4 is constructed.The integrated verification of the human-like decision-making and planning strategy based on cognition of traffic situation is carried out on the Driver-4 dataset.The verification results indicate that the proposed strategy achieves effective human-like cognition of the traffic situation,can generate driving behaviors and trajectories that conform to the cognitive and behavioral characteristics of human drivers,and has good generalization.The proposed strategy can give the intelligent vehicle human-like driving ability,and outperforms human drivers in some decision-making and planning performances.