【作者】 赵伟；

【导师】 郭洪艳；

【作者基本信息】 吉林大学 ， 控制工程（专业学位）， 2022， 硕士

【摘要】 汽车工业的发展给人们的生活带来了极大的便捷,但是,在日常的生活中,由驾驶员因素导致的交通事故也屡见不鲜。随着计算机技术的发展,汽车的智能驾驶技术引起了人们的广泛关注。路径规划作为智能驾驶技术中关键的一环,决定着智能驾驶系统的可靠性与安全性。面对简单的驾驶场景,车辆要做出合理的路径规划,这是对智能系统最基本的要求,但在复杂多变的交通环境中,智能车辆也应该做出相应的操作躲避危险,既要保证车辆的驾驶安全,也要保证车辆的驾驶稳定性,这给路径规划带来了更高的要求。恶劣的行车工况也会给路径规划的安全性带来挑战,如冰雪环境下车辆容易发生侧滑,所以,如何在冰雪环境下规划出一条安全的可行驶路径是另一个研究难点。针对上述路径规划技术中的难点,本文展开了对智能车辆自主避障路径规划的研究:针对路径规划安全性和稳定性的问题,提出了基于Q学习的换道超车路径规划方法。选用合适的参数对车辆的状态和动作进行定义,简化了规划问题,减少了车辆状态的数量,提高了算法的效率;设计了兼顾车辆安全性、稳定性和路径平滑性的奖励函数,实现了车辆换道超车的驾驶目标。仿真结果表明,虽然由于路径的平滑性欠佳导致车辆在执行过程中存在一些偏差,但在一定的误差范围内保证了行车的安全性。针对规划路径的平滑性不佳导致车辆执行存在偏差的问题,提出了考虑车辆动力学的深度Q学习路径规划方法。由于车辆动力学相对复杂,车辆的状态参量增加,使用Q学习算法会出现维度灾难的问题,因此,采用基于神经网络的深度Q学习算法,改善了Q学习算法中Q表存储量有限的问题;考虑道路环境相对复杂,采用人工势场的方法,建立了包含车道中心线、障碍车辆和驾驶目的地的道路环境虚拟力场描述,引入了相似状态判断机制,使车辆在面对没有训练过的对称状态时,也能做出相应的对称动作,适应了更多的交通场景。仿真结果表明,该方法增加了所规划路径的平滑程度,减小了车辆在行驶过程中侧向参数的峰值,提高了车辆的侧向稳定性,减少了深度Q学习的训练时间,降低了学习成本。针对冰雪环境工况路面附着系数低,导致车辆易出现侧滑现象的问题,本文提出一种考虑周围障碍车辆发生侧滑的路径规划方法。分析了冰雪环境对车辆驾驶安全的影响,考虑到车辆发生侧滑现象时的危险程度更高,改进了在前文提出的路径规划算法,扩展了障碍车动态矩形虚拟斥力场的作用范围,增强斥力场的斥力,达到了使车辆及时避开危险车辆、避免碰撞的发生的目标。仿真结果表明,无论车辆是在车道线上正常行驶,还是为了躲避障碍车辆进行换道操作,车辆都能及时规划出安全的路径,避免碰撞事故的发生。更多还原

【Abstract】 The development of automobile industry has brought great convenience to people’s life.However,in daily life,traffic accidents caused by driver factors are also common.With the development of computer technology,the intelligent driving technology of automobile has attracted extensive attention.As a key part of intelligent driving technology,path planning determines the reliability and safety of intelligent driving system.In the face of simple driving scenes,the vehicle should make reasonable path planning,which is the most basic requirement for the intelligent system.However,in the complex and changeable traffic environment,the intelligent vehicle should also make corresponding operations to avoid danger.It is necessary to ensure the driving safety and driving stability of the vehicle,which brings higher requirements to the path planning.Bad driving conditions will also bring challenges to the safety of path planning.For example,vehicles are prone to sideslip in the ice and snow environment.Therefore,how to plan a safe driveable path in the ice and snow environment is another research difficulty.In view of the difficulties in the above path planning technology,this paper studies the path planning of autonomous obstacle avoidance of Intelligent Vehicles:Aiming at the problem of safety and stability of path planning,a lane changing overtaking path planning method based on Q-Learning is proposed.Selecting appropriate parameters to define the vehicle state and action simplifies the planning problem,reduces the number of vehicle states and improves the efficiency of the algorithm;A reward function considering vehicle safety,stability and path smoothness is designed to achieve the driving goal of vehicle lane changing and overtaking.The simulation results show that although there are some deviations in the implementation process due to the poor smoothness of the path,the driving safety is guaranteed within a certain error range.Aiming at the problem that the poor smoothness of the planned path leads to the deviation of vehicle execution,a deep Q-learning path planning method considering vehicle dynamics is proposed.Due to the complexity of vehicle dynamics and the increase of vehicle state parameters,the use of Q-learning algorithm will lead to the problem of dimensional disaster.Therefore,the deep Q-learning algorithm based on neural network is adopted to improve the problem of limited q-table storage in Q-learning algorithm;Considering the relatively complex road environment,using the method of artificial potential field,a virtual force field description of road environment including Lane centerline,obstacle vehicles and driving destination is established,and a similar state judgment mechanism is introduced,so that vehicles can make corresponding symmetrical actions in the face of untrained symmetrical States,so as to adapt to more traffic scenes.The simulation results show that this method increases the smoothness of the planned path,reduces the peak value of lateral parameters during driving,improves the lateral stability of the vehicle,reduces the training time of deep Q-learning and reduces the learning cost.Aiming at the problem that the road adhesion coefficient is low under the condition of ice and snow,which leads to the phenomenon of vehicle sideslip,this paper proposes a path planning method considering the sideslip of surrounding obstacle vehicles.The influence of ice and snow environment on vehicle driving safety is analyzed.Considering that the danger degree of vehicle sideslip is higher,the path planning algorithm proposed above is improved,the action range of obstacle vehicle dynamic rectangular virtual repulsion field is expanded,and the repulsion of repulsion field is enhanced,so as to achieve the goal of avoiding dangerous vehicles and collision in time.The simulation results show that whether the vehicle is driving normally on the lane line or changing lanes to avoid obstacles,the vehicle can plan a safe path in time to avoid collision accidents.更多还原