基于强化学习优化汽车智能换道与间距预警的汽车避障系统  

Vehicle obstacle avoidance system based on reinforcement learning and optimization of intelligent lane change and gap warning

作　　者：张立霞[1] 张莉莉[1] 刘晋丽 ZHANG Lixia;ZHANG Lili;LIU Jinli(Xi’an Aeronautical Polytechninc Institute,Xi’an 710089,China)

机构地区：[1]西安航空职业技术学院,西安710089

出　　处：《自动化与仪器仪表》2023年第8期139-143,共5页Automation & Instrumentation

基　　金：西安航空职业技术学院2020年度科研计划项目《汽车避障伴随辅助驾驶系统的设计》(20XHZK-06)。

摘　　要：高速行驶的汽车在道路换道时需要保持安全的间距与车速,为保障汽车在换道过程安全行驶,提出一种基于动态空间转换法(Ego Dynamic Space Transform,EDST)与强化学习(Double Deep Q Network,DDQN)的多场景汽车避障预警算法。将单目深度预估图作为汽车航点最佳时刻,采用DDQN算法检测图像输入并执行动作输出。由于车辆换道场景的复杂性,采用对抗学习法(Adversarial Discriminative Domain Adaptation,ADDA)处理目标场景数据,实现车辆不同场景下的换道操作。选择多种场景测试车辆避障模型性能,所提出的自适应模型在复杂双向车道场景碰撞次数最少为3次。同时,能够换道数量最多为42次,优于EDST、DDQN以及DDQN+EDST模型,满足智能汽车安全换道要求。研究内容为高速驾驶车辆紧急避险提供重要的技术参考。High-speed vehicles need to maintain a safe distance and speed when changing lanes.In order to ensure the safety of vehicles in the process of changing lanes,a multi-scene vehicle obstacle avoidance early warning algorithm based on Ego Dynamic Space Transform(EDST)and Double Deep Q Network(DDQN)is proposed.The monocular depth prediction map is taken as the best time of the vehicle waypoint,and the DDQN algorithm is used to detect the image input and execute the action output.Due to the complexity of vehicle lane change scenarios,the adversarial discriminative domain adaptation(ADDA)method is used to process the target scene data to achieve vehicle lane change operations in different scenarios.Multiple scenarios are selected to test the performance of the vehicle obstacle avoidance model.The proposed adaptive model has at least 3 collisions in the complex two-way lane scene.At the same time,it can change lanes up to 42 times,which is superior to the EDST,DDQN and DDQN+EDST models,and meets the requirements for safe lane change of intelligent vehicles.The research content provides important technical reference for emergency avoidance of high-speed vehicles.

关 键 词：DDQN EDST ADDA 智能换道 预警 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置] U471.15[自动化与计算机技术—控制科学与工程]