Research On Key Technologies Of Human-like Decision Making System For Autonomous Vehicles

Autonomous driving technology has great potential to solve the mobility challenges brought by the increasing demand for driving safety,traffic efficiency,energy saving and environmental protection.The ability of autonomous decision-making is the core embodiment of the intelligence of autonomous driving.Existing decision making systems are mainly designed for low-level autonomous vehicles and are not qualified for application of high-level autonomous driving for their disadvantages,such as insufficient real-time trajectory planning,insufficient prediction function for surrounding vehicles,and limited adaptability in different traffic conditions.In a word,the overall performance is obviously below the level of human drivers.This study focuses on decision making system for autonomous vehicles and provides solutions to the key technologies such as the online spatial-temporal trajectory planning,the joint prediction of the lane changing intention and trajectory,and the design of the human-like reward function for autonomous driving,which leas to the realization of human-like decision-making function for autonomous vehicles under highway scenario.Firstly,a spatiotemporal trajectory planning method based on directed acyclic graph is proposed.Aiming at dynamic traffic flow and road constraints,a directed acyclic map with time information is constructed based on grid map to express environmental prediction information,and an improved A* algorithm is used to achieve synchronous search of path and speed.Furthermore,a convex feasible region for autonomous driving is constructed based on surrounding dynamic obstacles.The aforementioned spatiotemporal trajectory is then optimized by the model predictive control method combined with the vehicle kinematics model,which realizes the online spatiotemporal trajectory planning for autonomous vehicles.Secondly,a surrounding vehicle trajectory prediction method based on lane change intention recognition is proposed.In view of environmental uncertainty and sensor noise,different statistical indicators are combined to generate candidate features based on synchronous sliding window technology and conditional mutual information index is used to optimize these features.Based on random forest classification probability output,D-S evidence theory is introduced for multi-source determination of lane-changing intention.Furthermore,with help of long short term memory network,a unified trajectory prediction neural network architecture with semantic information of lane-changing intention is designed to achieve long-term trajectory prediction for surrounding vehicles.Finally,an inverse reinforcement learning method based on pre-sampled driving strategy is proposed.In view of the inefficiency of inverse reinforcement learning of driver’s decision-making reward function,qualitative analysis points out that to avoid solving reinforcement learning problem is a powerful way to accelerate learning.Furthermore,the driver model knowledge is used to construct the driving strategy subspace,and the candidate trajectories are generated based on strategy pre-sampling.Then the trajectory optimization mechanism is designed and the optimal weight coefficients are learned iteratively by gradient updating,which realizes efficient learning of the decision-making reward function.To avoid heavy cost and high risk of verifying autonomous driving system on real road,an autonomous driving system simulation platform is developed.The proposed human-like decision making method has been verified in terms of driving safety,time efficiency,ride comfort and fuel economy,which has the characteristics of result optimality,real-time computation and platform mobility.This study proposes a systematic research framework and method for the research of human-like decision making system for autonomous vehicles,and promotes the interdisciplinary study between knowledge-driven problem modeling and data-driven machine learning,which provides theoretical guidance and technical support for the practical application of high-level autonomous driving.