Semantic Grid Map based LiDAR Localization in Highly Dynamic Urban - - PowerPoint PPT Presentation
Semantic Grid Map based LiDAR Localization in Highly Dynamic Urban Scenarios 12 th IROS20 Workshop on Planning, Perception and Navigation for Intelligent Vehicles Oct. 2020 Chenxi Yang, Lei He, Hanyang Zhuang, Chunxiang Wang, Ming Yang *
▪ Chenxi Yang, Lei He, Hanyang Zhuang, ▪ Chunxiang Wang, Ming Yang* ▪ * mingyang@sjtu.edu.cn ▪ This work is supported by the National Natural Science Foundation of China (U1764264/61873165)
12th IROS20 Workshop on Planning, Perception and Navigation for Intelligent Vehicles
Contents 1
2 3
4
5
Signal denial Multipath effect
Robust to illuminance High reliability
Dynam namic ic interferen rferences ces
Illuminance changes Low reliability
Contents 1 2 3 4 5
GNSS-based sed
✓ global consistency
SLAM-base sed
✓ local consistency
Point nt cloud
✓ accuracy
2D 2D gr grid d map
✓ data size & speed
Feature re map
✓ accuracy & speed
LiDAR-based localization Non- semantics Point
IMLS-SLAM (ICRA2018)
Grid
Probabilistic Maps (ICRA2010) Robust Localization (ICRA2017)
Feature
LOAM (ICRA2014) SuMa (ICRA2018）
Descriptor
Segmatch (RSS2018) L3-Net (CVPR2019)
Semantics Semantic
feature
Lane (IROS2018) Pole-like (IROS2016）
Point cloud segmentation
Semantic ICP (BMVC2018)
Represent esentati ative ve methods
Non-real time Dynamic interference Dynamic interference Dynamic interference
Feature missing Robustness Non-real time
Non-semantics: dynamic interference Semantics: difficult to find a balance between real-time and robustness
MINES Stanford Baidu CMU Bonn KIT Freie
UMich
Agency cy
ETH Baidu
Non-real time Dynamic interference
Challe lenges ges
Contents 1 2 3 4 5
▪ Featur ture e selecti ection
▪ Abundant in urban scenarios ▪ Strongly imply static ▪ Extractable from scan-level sparse point cloud ▪ Sufficient pose constraints from multiple layers
▪ Seman antic tic grid d map
▪ To speed up the calculation ▪ Semantic category with a trust rate
Contents 1 2 3 4 5
▪ On On-li line ne pose se initiali ializa zation tion
▪ Large range search ▪ Limited to the first several frames ▪ Relatively low real-time requirements
▪ Real-tim time e traje jecto ctory ry tracki cking ng
▪ Can inherit an accurate initial position from the previous frame ▪ Every frame ▪ Strict real-time requirements (typically 100ms)
▪ On On-li line ne pose se initiali ializa zation tion
▪ Notation Map Cubes Scan Cubes
Semantic Category
GMM
▪ Real-tim time traje jecto ctory ry tracki cking
▪ Notation ▪ Residual error
1 2 3 4 5
Contents
▪ Proces cesso sor
▪ Intel i7-7567U @3.5GHz with 16GB memory
▪ Express press road
▪ 5.2km long
▪ On On-li line ne pose se initiali ializa zation tion
▪ (0.2m) 3 cube ▪ horizontal offset uniform distribution in 50m circle ▪ up to 90 degree offset ▪ a s speci cial al case Conjuncti nction
Initial ial position ion CPD Ours 2nd
d iterat
ation
Result lt
▪ Real-tim time traje jecto ctory ry tracki cking
▪ (0.1m) 2 square
▪ Proces cesso sor
▪ Intel i7-7567U @3.5GHz with 16GB memory
▪ Facto ctory
▪ 1.5km long