融合车辆轨迹预测的学习型自动驾驶决策  

Learning-based automatic driving decision-making integrated with vehicle trajectory prediction

作　　者：徐杰 裴晓飞[1,2] 杨波[1] 方志刚[2] XU Jie;PEI Xiaofei;YANG Bo;FANG Zhigang(Hubei Key Laboratory of Advanced Technology of Automotive Components,Wuhan University of Technology,Wuhan 430070,China;Hubei Collaborative Innovation Center of Automotive Components Technology,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]现代汽车零部件技术湖北省重点实验室,武汉理工大学,湖北武汉430070 [2]汽车零部件技术湖北省协同创新中心,武汉理工大学,湖北武汉430070

出　　处：《汽车安全与节能学报》2022年第2期317-324,共8页Journal of Automotive Safety and Energy

基　　金：湖北省技术创新重大专项(2020DEB014);现代汽车零部件技术湖北省重点实验室开放基金(XDQCKF2021009)。

摘　　要：在考虑车辆未来轨迹的基础上,利用强化学习来实现在复杂场景下的驾驶决策问题。基于图结构和长短时记忆(LSTM)网络搭建周围车辆长期交互的轨迹预测模型,利用Rainbow DQN算法构建行为决策模型。在该模型中,状态空间包括当前时刻自车及旁车信息和未来的轨迹预测信息;从安全性、舒适性、行车效率等角度设计对应的奖励函数,并设定安全规则提高所选动作的安全性。结果表明:第5 s末车辆轨迹预测纵向位置误差为1.54 m,横向位置误差为0.32 m,效果较为准确。因而,加入轨迹预测有利于提高自动驾驶汽车决策的安全性和通行效率。On the basis of considering the future trajectory of the vehicle,reinforcement learning was used to realize the decision-making problem of driving in a complex scenario.A long-term interaction trajectory prediction model of surrounding vehicles was built based on the graph structure and Long Short Term Memory(LSTM)and Rainbow DQN algorithm was used to build a behavioral decision model.In this model,the state space not only considered the current time of the vehicle information,but also considered the future trajectories of these vehicles.The corresponding reward function was designed from the perspectives of safety,comfort,driving efficiency,etc.Safety rules were set to improve the safety of selected actions.The results show that at the end of 5 s,the method with vehicle trajectory prediction has a longitudinal location error of 1.54 m with a lateral location error of 0.32 m,which are relatively accurate.Therefore,this method improves the safety and efficiency of decision-making for autonomous vehicles.

关 键 词：汽车工程 自动驾驶 强化学习 决策模型 车辆轨迹预测 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]