深度学习在自动驾驶领域应用综述  

A Review of Deep Learning Applications for Autonomous Driving

作　　者：段续庭 周宇康 田大新 郑坤贤 周建山 孙亚夫 DUAN Xuting;ZHOU Yukang;TIAN Daxin;ZHENG Kunxian;ZHOU Jianshan;Sun Yafu(Beijing Key Laboratory for Cooperative Vehicle Infrastructure Systems and Safety Control,School of Transportation Science and Engineering,Beihang University,Beijing 102206,China;National Engineering Laboratory for Comprehensive Transportation Big Data Application Technology(NEL-CTBD),Beijing 100191,China;China TRANSINFO,Beijing 100085,China)

机构地区：[1]北京航空航天大学交通科学与工程学院车路协同与安全控制北京市重点实验室,北京102206 [2]综合交通大数据应用技术国家工程实验室,北京100191 [3]北京千方科技股份有限公司,北京100085

出　　处：《无人系统技术》2021年第6期1-27,共27页Unmanned Systems Technology

基　　金：国家自然科学基金资助项目(62173012,U20A20155,61822101);北京市自然科学基金资助项目(L191001);国家自然科学基金中英牛顿高级学者计划(62061130221)。

摘　　要：成熟的自动驾驶技术能够极大降低交通事故率,保障驾驶人员与行人的安全,优化交通流运行,但早期的自动驾驶系统可靠性与智能性都很低,不能满足实际应用需求。近年来,深度学习技术迅速发展,并与自动驾驶领域结合,其在机器视觉、自然语言处理等领域的成功应用使得自动驾驶越来越接近现实。介绍了目前自动驾驶系统的主流技术框架,并对其各模块中深度学习技术的应用情况进行系统梳理,将自动驾驶系统分为分解式和端到端式两种技术方案,并将分解式方案进一步分为感知、决策、控制3大模块,分别对以上两类解决方案中深度学习技术应用的历史沿革、研究现状以及典型算法性能进行综合评述。已有的研究成果表明,分解式方案的技术路径较为成熟,感知、决策、控制3个功能模块分工清晰,可解释性强,但系统复杂度高,计算量大,软件架构庞大,硬件要求高,应进一步简化各问题的算法,加强各个算法模块间的功能整合,降低系统复杂度与硬件要求;端到端式方案计算量小,硬件要求低,且系统复杂度低,但对算法要求高,安全性低,可解释性、可靠性差,建议未来通过完善智能道路基础设施,推进5G传输的应用,加强车路、车云协同,进一步完善现有算法来解决以上问题。The rate of traffic accidents can be reduced by mature autonomous driving technology.The technology can guarantee the safety of drivers and pedestrians,and also optimize the operation of traffic flow.However,the reliabil-ity and intelligence of early autonomous driving systems were inadequate,which could not meet practical requirements.In recent years,thanks to the rapid development of the deep learning technology,the autonomous driving technology is much closer to reality with its successful applications in Machine Vision,Nature Language Processing and other fields.This article introduces some dominant technical frameworks in autonomous driving systems,and conduct a systematic analysis in the application of different modules of deep learning technology.The autonomous driving system can be di-vided into two technical plans,decomposition plan and end-to-end plan.The decomposition plan can be divided further into three modules,perception,decision-making,and control.The article conducts a comprehensive assessment of the evolution histories,contemporary researches and performance of typical deep learning algorithm in autonomous driving systems.Existing research results show that the technique of decomposition plan is relatively more mature,with a clearer distinguishment of functions among three modules,as well as a better interpretability.However,the decomposi-tion system is complex and requires high computing capability large software structure and cutting-edge hardware.The algorithm should be simplified,the integration of functions between each algorithm module should be enhanced,and the complexity and hardware requirements should be decreased.On the other hand,the end-to-end plan has a smaller com-puting ability requirements and does not require much on hardware,while it requires a better algorithm and the safety level,interpretability and reliability is rather low.The problems above are suggested to be solved by implementing intel-ligent road infrastructure,the applications of 5G transmission technology,the stren

关 键 词：深度学习 自动驾驶 目标检测 路径规划 端到端学习 感知 决策 控制 

分 类 号：U471.15[机械工程—车辆工程]