| Highway driving is a challenging task.It requires intelligent decision-making to achieve long-term goals and short-term trajectory planning to safely execute decisions.In the decision planning on automatic driving,few car manufacturers have studied a complete module.Most of the decision-making module generates some actions based on the sensors and outputs this action to the planning module for a safe and comfortable trajectory.Numerous works examines an integral module but considering different application scenarios and different development ways of the module.In many of the existing research methods,there are a variety of contingency mechanisms to deal with emergencies when faced with different road environments,and there is rarely a more comprehensive approach to dealing with them.In this dissertation,we propose to use popular artificial intelligence and some traditional theoretical methods based on mathematical to solve this problem and propose the decision on how to design the planning module in this dissertation.The method refers to previous research ideas to complete the design of the decision planning module with task completion in mind for different driving tasks and a variety of difficult environments.The main studies are as follows.(1)Firstly,traditional behavioral decision-making methods are used.This study proposes a hierarchical structure of decision-making and planning for the road driving task.This dissertation uses intelligent driving models(IDM and MOBIL)to make long-term decisions based on traffic conditions.These decisions maximize the self-study performance while respecting the goals of other vehicles;secondly,a combination of deep learning and reinforcement learning is used.This traditional decision-making approach is used as a baseline method when evaluating deep reinforcement learning approaches.The deep reinforcement learning algorithm used in this dissertation takes the vehicle’s trajectory state,images,and information obtained from the vehicle’s radar and IMU as inputs for the decision-making part of the self-driving car.(2)In the face of complex traffic scenarios,this dissertation combines polynomial-based planning algorithms and numerical optimization algorithms for trajectory planning.Since there be various unexpected conditions in a complex environment,the numerical optimization has a combination of collision and comfortability to design the trajectory optimization so that it can cope with a variety of driving conditions.(3)Use the current popular deep learning framework to build the Conditional Imitation Learning(CIL)network in this dissertation,which is mainly used for images to carry out deep information mining,so that it can train an "excellent" model.This network is mainly used for deep information mining of images so that a better model can be trained,which can perform accurate decisions when combined with reinforcement learning algorithms.(4)Finally,it is necessary to validate the decision-planning system of this study.To compare and analyze the output of the different algorithms designed in this study,the dissertation show that the decision-planning system of is effective and can perform various tasks robustly. |