欢迎访问《汽车安全与节能学报》,

JASE ›› 2019, Vol. 10 ›› Issue (2): 119-145.DOI: 10.3969/j.issn.1674-8484.2019.02.001

• 综述与展望 •    下一篇

深度神经网络的关键技术及其在自动驾驶领域的应用

李升波 1,关  阳 1,侯  廉 1,高洪波 1,段京良 2,梁  爽 3,汪  玉 3,成  波 1, 李克强 1,任  伟 4,李  骏 1#br#   

  1. (1. 清华大学 车辆与运载学院,北京100084,中国;2. 加州大学伯克利分校 机械系,加州  94720,美国; 3. 清华大学 电子工程系,北京100084,中国;4. 加州大学河滨分校 电子计算机系,加州92521,美国)
  • 收稿日期:2019-01-19 出版日期:2019-06-29 发布日期:2019-07-05
  • 作者简介:李升波(1982—),男(汉),山东,副教授。E-mail: lishbo@tsinghua.edu.cn。
  • 基金资助:

    “十三五”国家重点研发计划(2016YFB0100906);国家自然科学基金面上项目(51575293);国家自然科学基 金优秀青年科学基金项目(U1664263);国家自然科学基金重点项目(51622504);北京市自然科学基金杰出青 年科学基金项目(JQ18010);汽车安全与节能国家重点实验室开放基金课题(KF1828)。

Key technique of deep neural network and its applications in autonomous driving

LI Shengbo1, GUAN Yang1, HOU Lian1, GAO Hongbo1, DUAN Jingliang2, LIANG Shuang3, WANG Yu3, CHENG Bo1, LI Keqiang1, REN Wei4, LI Jun1   

  1. (1. School of Vehicle and Mobility, Tsinghua University, Beijing 100084, China; 2. Mechanical Engineering, University of California Berkeley, Berkeley, CA 94720, USA; 3. Electronic Engineering, Tsinghua University, Beijing 100084, China; 4. Electrical and Computer Engineering, University of California Riverside, Riverside, CA 92521, USA
  • Received:2019-01-19 Online:2019-06-29 Published:2019-07-05

摘要:

        智能化是汽车的三大变革技术之一,深度学习具有拟合能力优、表征能力强和适用范围广的 特点,是进一步提升汽车智能性的重要途径。该文系统性总结了用于自动驾驶汽车的深度神经网络 (DNN)技术,包括发展历史、主流算法以及感知、决策与控制技术应用。 回顾了神经网络的历史及现状, 总结DNN的“神经元-层-网络”3级结构,重点介绍卷积网络和循环网络的特点以及代表性模型; 阐述了以反向传播(BP)为核心的深度网络训练算法,列举用于深度学习的常用数据集与开源框架,概 括了网络计算平台和模型优化设计技术;讨论DNN在自动驾驶汽车的环境感知、自主决策和运动控 制3大方向的应用现状及其优缺点,具体包括物体检测和语义分割、分层式和端到端决策、汽车纵 横向运动控制等;针对用于自动驾驶汽车的DNN技术,指明了不同问题的适用方法以及关键问题的 未来发展方向。

关键词: 智能汽车 , 自动驾驶 , 深度神经网络(DNN) , 深度学习 , 环境感知, 自主决策 , 运动控制

Abstract:

Autonomous driving is one of the three major innovations in automotive industry. Deep learning is a crucial method to improve automotive intelligence due to its outstanding abilities of data fitting, feature representation and model generalization. This paper reviewed the technologies of deep neural network (DNN) for autonomous vehicles, which covered its history, main algorithms and key technical application. The historical timeline of DNN, its “Unit-Layer-Network” architecture, and two types of representative models were introduced. The training algorithms centered on back propagation (BP), labelled datasets and free-source frameworks for deep learning were summarized, followed by the introduction to computing platforms and model optimization technologies. Finally, the applications of DNN in autonomous vehicles were discussed, including object detection and semantic segmentation, hierarchical and end-to-end decision-making, longitudinal and lateral motion control. The applicable methods and future works for different key problems of DNN in autonomous vehicles were pointed out.

Key words: intelligent vehicles , autonomous driving , deep neural network(DNN) ,  deep learning ,  environmental perception , decision making , motion control