汽车工程 ›› 2023, Vol. 45 ›› Issue (12): 2310-2317.doi: 10.19562/j.chinasae.qcgc.2023.12.013

所属专题: 智能网联汽车技术专题-控制2023年

• • 上一篇    下一篇

基于GCN和CIL的端到端自动驾驶换道方法

吕彦直,魏超(),何元浩   

  1. 北京理工大学机械与车辆学院,北京 100081
  • 收稿日期:2023-04-11 修回日期:2023-05-29 出版日期:2023-12-25 发布日期:2023-12-21
  • 通讯作者: 魏超 E-mail:weichaobit@163.com
  • 基金资助:
    国家自然科学基金(U1764257)

An End-to-End Lane Change Method for Autonomous Driving Based on GCN and CIL

Lü Yanzhi,Chao Wei(),Yuanhao He   

  1. School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081
  • Received:2023-04-11 Revised:2023-05-29 Online:2023-12-25 Published:2023-12-21
  • Contact: Chao Wei E-mail:weichaobit@163.com

摘要:

本文针对自动驾驶的换道行为,为了解决传统的端到端方法存在输出不稳定、动态交互场景特征信息难以提取的问题,提出了一种基于图卷积网络和条件模仿学习的自主换道端到端学习方法。首先以图结构数据的形式对驾驶场景的动态交互信息进行聚合,通过图卷积网络输出自车应采取的驾驶行为指令;然后与条件模仿学习结合,图卷积网络输出的驾驶指令作为指导条件模仿学习的高级命令,结合其他感知数据最终映射到车辆的底层控制动作,完成无碰撞自主换道;最后在CARLA仿真平台进行了实验验证。实验结果表明:所提出方法的性能优于传统的端到端方法,且具有更好的实验成功率以及泛化性能。

关键词: 智能车辆, 端到端驾驶, 换道, 条件模仿学习, 图卷积网络

Abstract:

For the lane change of autonomous driving, to solve the problems of unstable output and difficulty to extract dynamic interactive scene feature in conventional end-to-end method, an end-to-end learning method for autonomous lane change based on graph convolutional network (GCN) and conditional imitation learning (CIL) is proposed in this paper. Firstly, the dynamic interactive information of driving scenarios is aggregated in the form of graph-structured data. Secondly, the driving behavior instructions that the ego vehicle should take are output through GCN, which is then combined with CIL. The driving instructions output by GCN are taken as high-level commands for guiding CIL, and are finally mapped to underlying control actions of the vehicle with other perception data to complete autonomous lane change without collision. Experimental verification is carried out on CARLA simulation platform. The experimental results prove that the performance of this method is better than that of conventional end-to-end method, and it has better success rate and generalization performance.

Key words: intelligent vehicle, end-to-end driving, lane change, conditional imitation learning, graph convolutional network