Civil Maps是一家三维地图初创公司,总部位于加州。CivilMaps 从汽车内部传感器收集数据,并将其转化成为“智能地图信息”供自动驾驶汽车使用。本文原作者为CivilMaps,佐思产研做了简单翻译整理。
车辆认知:从传感器、认知引擎、决策引擎、到车载控制
![](/UpLoads/IndustryNews/2018_05/3a5c6344-470e-4c67-af55-a5955f161cb8.jpg)
车辆认知:从传感器到本地缓存(包括映射、定位及假设检验过程)
![](/UpLoads/IndustryNews/2018_05/bd048a63-0fb0-4a60-ae6e-ce21d6fe9303.jpg)
Civil Maps管理团队
![](/UpLoads/IndustryNews/2018_05/e229dc52-99c3-4e4c-b4ff-16c48f993bb8.jpg)
无人驾驶汽车传感器(激光雷达、摄像头、OBD、GPS/IMU)
![](/UpLoads/IndustryNews/2018_05/737a5732-8cfb-4338-93b3-de85a0df296a.jpg)
传感数据到驾驶决策
![](/UpLoads/IndustryNews/2018_05/81891b04-d96e-4a0a-9455-b8d16852744e.jpg)
认知层
![](/UpLoads/IndustryNews/2018_05/f8042f17-84dd-4144-b082-924ff3c88f3e.jpg)
认知层:秘籍
![](/UpLoads/IndustryNews/2018_05/7badfd82-0781-4295-8aaa-28b80e77cc46.jpg)
Atlas DevKit 系统
![](/UpLoads/IndustryNews/2018_05/51755112-a4df-421a-8c5b-93918580ffaa.jpg)
传感器融合的必备条件
![](/UpLoads/IndustryNews/2018_05/dbcaba9c-dcc5-4575-9e48-cd5b5a049b16.jpg)
Civil Maps系统
![](/UpLoads/IndustryNews/2018_05/10f89c7c-7488-4146-8db2-ecb44fefd6e9.jpg)
惯性测量单元传感器融合(一阶)
![](/UpLoads/IndustryNews/2018_05/26ee1f95-cccc-4810-a8bf-317ee480d724.jpg)
激光雷达测程法(一阶)
![](/UpLoads/IndustryNews/2018_05/e560dc2c-ccda-483c-a1df-63feb16c0495.jpg)
摄像测程法(一阶)
![](/UpLoads/IndustryNews/2018_05/3ca60894-2ac9-444a-afcd-bbe03e6bf6ac.jpg)
总结:一阶传感器融合
![](/UpLoads/IndustryNews/2018_05/1ff7566a-4319-49d1-9a91-fb91869dcfc2.jpg)
二阶传感器融合
摄像头投射(二阶融合)(RGB FoV + LiDAR + SBET) = RGB PC
![](/UpLoads/IndustryNews/2018_05/d012d6d8-97ca-4388-904f-631aac97a8fa.jpg)
三阶传感器融合
难点1:时走时停交通路段(stop-and-go, S&G),IMU偏移会使参考地图失真
![](/UpLoads/IndustryNews/2018_05/a418f44b-ab41-4d26-b37d-2ad4ad1061bb.jpg)
最佳平滑轨迹估算(SBET)校准(三阶融合)
(RGB PC + 能量) = SBET校准
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626669356805829.jpg)
通过能量函数以及OBD对IMU速度向量进行校准
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626669475704469.jpg)
众包地图:客户端/服务器
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626669578344384.jpg)
认知层:堆栈管道
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626669808358959.jpg)
认知层
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626669871606864.jpg)
地理空间指纹技术
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626669982927864.jpg)
指纹基础地图:自动驾驶车辆如果知道自己的位置
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670143396109.jpg)
利用指纹技术,将基础点云地图进行3500倍压缩,从14G压缩到47MB
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670238145904.jpg)
指纹技术能够降低计算和功耗
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670323980254.jpg)
运营成本降低300倍
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670417987909.jpg)
传统地理围栏vs. 游动边缘绘图
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670499691664.jpg)
游动vs传统测量
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670558047304.jpg)
基于边缘的众包
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670685783269.jpg)
3D语义地图
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670745252119.jpg)
增强现实地图+定位
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670803890944.jpg)
通过6个自由度的定位实现自动驾驶车辆认知
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670863203554.jpg)
没有增强现实,传感器需扫描整个场景后,才能识别交通信号灯状态
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626670931148424.jpg)
有增强现实,地图会通知车辆哪儿是优先注意点,减少认知负担
![](http://www.shujubang.com/UpLoads/Special/File/2018_05/636626671008955944.jpg)