用于自动驾驶汽车的深度学习技术介绍  

Introduction to Deep Learning and its Application on Autonomous Vehicles

作　　者：李升波[1] 张航 Li Shengbo;Zhang Hang(School of Vehicle and Mobility,Tsinghua University,Beijing,100084;School of Computer and Information Engineering,Tianjin Agricultural University,Tianjin,300384)

机构地区：[1]清华大学车辆与运载学院,北京100084 [2]天津农学院计算机与信息工程学院,天津300384

出　　处：《建设科技》2022年第1期37-46,共10页Construction Science and Technology

基　　金：“十三五”国家重点研发计划(2016YFB0100906);北京市自然科学基金资助(杰青:JQ18010);汽车安全与节能国家重点实验室开放基金课题(KF1828)支持。

摘　　要：智能化是汽车的三大变革技术之一。深度学习(Deep Learning,DL)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。本文总结了用于自动驾驶汽车的深度学习技术,包括发展历史、主流算法以及感知、决策与控制技术应用。首先回顾深度学习的历史及现状,总结神经网络的“神经元-层-网络”三级结构,重点介绍卷积网络和循环网络的特点以及代表性模型。其次阐述以反向传播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,概括网络计算平台和模型优化设计技术。最后讨论深度学习在自动驾驶汽车的环境感知、自主决策和运动控制三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、分解式和端到端决策、汽车纵横向运动控制等,针对用于自动驾驶汽车的深度学习技术,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。Autonomous driving is one of the three major innovations in automotive industry.Deep learning(DL)is a crucial method to improve automotive intelligence due to its outstanding abilities of data fitting,feature representation and model generalization.This paper reviews the DL technologies for autonomous vehicles,which covers its history,main algorithms and key technical application.Firstly,the history and current situation of deep learning are reviewed,and the three-level structure“Unit-Layer-Network”of neural network is summarized,and the characteristics and representative models of convolutional network and recurrent network are emphatically introduced.Secondly,the training algorithms centered on back propagation is described.The labelled datasets and free-source frameworks for DL are listed,followed by the introduction to computing platforms and model optimization technologies.Finally,the applications as well as the advantages and disadvantages of DL in autonomous vehicles are discussed in three directions of autonomous perception,autonomous decision making and motion control of autonomous driving vehicle,including object detection and semantic segmentation,hierarchical and end-to-end decision-making,and vehicle longitudinal and lateral motion control.The applicable methods and future works for different key problems of DL in autonomous vehicles are also pointed out.

关 键 词：智能汽车 神经网络 深度学习 环境感知 自主决策 运动控制 

分 类 号：U463.6[机械工程—车辆工程] TP18[交通运输工程—载运工具运用工程]