笔记：深度学习驱动的自动驾驶新主流框架盘点

笔记：深度学习驱动的自动驾驶新主流框架盘点（附3篇论文）

提纲

• Nvidia DRIVE深度学习解决方案
• Mobileye深度学习方案
• Comma.ai深度学习方案

Nvidia DRIVE深度学习解决方案

• DGX-1 Training Platform
• DriveWorks PX2 传感器同步、标定、数据采集和记录
  英伟达论文下载：https://pan.baidu.com/s/1bo4jTvx

Mobileye深度学习方案 Shashua

Sensing: Environmental Model (360 awareness)

• 运动和静止的物体、车道线、可行驶区域和交通标志
• Moving/Stationary Objects:
  
  • 0 false-pos part of ADAS?
  • Vehicles at any angle
  • General (non-model-based) objects
• Path Delimiters: free-space, detect “any boundary”, label types of boundaries
• Drivable Paths
  
  • Detect lane boundaries of all drivable paths (even when lanes are non-existent)
    
    • host left/right, left/right mark of next lane left/right
  • Attach semantic meaning to each path (pavement markings and other contextual information)
  • Detect “key points” (merge, split, etc.)
• End-to-End vs Semantic Abstraction
  
  • 专家经验，领域知识，人工监督
  • End-to-End 对极端情况（Corner Cases）的应变能力比较差

高精度地图技术（Roadbook, REM，Road Experience Management）

• 纹理和色彩特征来生成地图
• 三维上稀疏，地面一维稠密的结构

三维元素

• 不会包括原始图像数据，而是经过识别后的语义信息
• Landmarks hierarchy:
  
  • Traffic signs (including arrows on the road and other marks) (~1000 tags, ~10bits)
  • Directional signs (~10000 tags, ~13bits)
  • General rectangular signs (signature ~50 bytes)
  • Lampposts and reflectors (<100tags)
  • Additional families of landmarks (e.g., dashed lines) will be added if needed

地面一维

• 道路模型，包括车道线的精确位置、连接关系等
• Location of lanes
• Stitch (ego-motion)
  
  • X(t) = a3 * t^3 + a2 * t^2 + a1 * t^1 + a0
  • Y(t) = b3 * t^3 + b2 * t^2 + b1 * t^1 + b0
  • Z(t) = c3 * t^3 + c2 * t^2 + c1 * t^1 + c0

Crowd Sourced

• Integration of REM segments (10kb per km)
• Increase accuracy of landmarks
• Change detection

Sensing vs Planning

• DeepMind采用的DQN网络不同，他们还考虑了驾驶过程中的时序性
• Reinforcement learning
• arXiv paper by Mobileye

Comma.ai深度学习方案

CNN + RNN

comma.ai 论文下载：https://pan.baidu.com/s/1pLPawVh

参考资料

【智驾深谈】深度学习驱动的自动驾驶新主流框架盘点（附3篇论文）
https://mp.weixin.qq.com/s?src=3&timestamp=1514026808&ver=1&signature=TuAMvlukldb23hcWJPdVOp8P8bAFnBRM-EHMq2o2BzSqNr0tcM7jsewvg-GR1COc9RsX9bZjiil8VNdcuTB8bgOug0ZcvtAyltEaOrE4DpV*GM8uTJ1KvTtiFFgbdW3o3kSEIhcyrBjUGQGXYLurTBTpt*qsJK6hl7xhZvslTjk=