ํ‚น๊ฐ“ Ji Zhang๋‹˜๊ป˜์„œ RSS 2014๋…„์— ๋ฐœํ‘œํ•˜์‹  LOAM ๋ฆฌ๋ทฐ
Mapping๋ณด๋‹ค๋Š” Odometry๋ฅผ ์ง‘์ค‘์ ์œผ๋กœ ๋ณด์ž!

LOAM ๊ฐœ์š”

๋…ผ๋ฌธ ์—ฐ๊ตฌํŒ€์€ odometry and mapping์„ ์œ„ํ•œ ์‹ค์‹œ๊ฐ„ method๋ฅผ ์ œ์•ˆ
Loop closure(LC)๊ฐ€ ์—†์ด low-drift & high accuracy ๋‹ฌ์„ฑ


ํ•˜๋“œ์›จ์–ด

img

2D ๋ผ์ด๋‹ค๋ฅผ ๋ถ€์ฐฉํ•˜์—ฌ ์‹œ๊ณ„๋ฐฉํ–ฅ/๋ฐ˜์‹œ๊ณ„๋ฐฉํ–ฅ์œผ๋กœ ํšŒ์ „
360๋„ ์Šค์บ”ํ•˜๋Š”๋ฐ 1์ดˆ ๊ฑธ๋ฆผ


LOAM ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ตฌ์กฐ

img

Odometry๋Š” 10Hz๋กœ ๋™์ž‘ ํ›„ Transformation ์ถœ๋ ฅ
Mapping์€ 1Hz๋กœ ๋™์ž‘ ํ›„ Transformation ์ถœ๋ ฅ
๊ฒฐ๊ณผ์ ์œผ๋กœ ๊ฐ ๋ชจ๋“ˆ์—์„œ ๋‚˜์˜จ Transformation์„ ํ†ตํ•ฉ!


Odometry Estimation

  1. Feature extraction

ํ•œ scan์„ 4๋ถ€๋ถ„์˜ subregion์œผ๋กœ ๋ถ„ํ• 
๊ฐ point์˜ ๊ณก๋ฅ ์„ ๊ณ„์‚ฐํ•จ

img

c๊ฐ’์„ ๊ธฐ์ค€์œผ๋กœ ๊ฐ subregion๋“ค์„ ์ •๋ ฌ์‹œํ‚ด!

c๊ฐ€ ํฌ๋ฉด Edge, ์ž‘์œผ๋ฉด Planar๋ผ๊ณ  ํŠน์ง• ๋ถ„๋ฅ˜

\(\therefore\) ์ด ํŠน์ง•์„ ๊ธฐ๋ฐ˜์œผ๋กœ Scan matching์„ ์ˆ˜ํ–‰ํ•  ๊ฑฐ๋‹ค!!
์ „ํ˜•์ ์ธ Keypoint ๊ธฐ๋ฐ˜ Scan matching, LOAM์—์„œ Keypoint๋Š” Edge์™€ Planar๊ฐ€ ๋จ


  1. Finding Feature Point Correspondence

๋งค์นญ์„ ์ˆ˜ํ–‰ํ•˜๋ ค๋ฉด Correspondence๋ฅผ ์ฐพ์•„์•ผ ํ•จ
๋ณ„๋‹ค๋ฅธ ๊ธฐ๋ฒ•์€ ์“ฐ์ด์ง€ ์•Š๊ณ  Euclidean Distance ๊ธฐ๋ฐ˜ ์ถ”์ถœ!
KDTree ์ž๋ฃŒ ๊ตฌ์กฐ๋ฅผ ์ด์šฉํ•ด์„œ ํ˜„์žฌ scene๊ณผ ์ง์ „ scene์˜ ํŠน์ง•์ [Edge/Planar]์˜ closest point์˜ ๊ฑฐ๋ฆฌ๊ฐ’์„ ๋น„๊ตํ•ด์„œ ํŠน์ง•์  ์ถ”์ถœ!!



์œ„์˜ ๋‘ ๊ฐœ๋ฅผ vectorize ํ•ด์„œ ์ตœ์ ํ™” ํ• ๊ฑฐ์ž„


  1. Motion estimation

๋ผ์ด๋‹ค๋Š” ์„ ํ˜• ์›€์ง์ž„์„ ๊ฐ€์ •ํ•˜๊ณ  ๋ชจ๋ธ๋ง
๋”ฐ๋ผ์„œ ๋ผ์ด๋‹ค ๋ชจ์…˜์˜ Linear interpolation์ด ๊ฐ€๋Šฅ

img

Translation์˜ ๊ณต์‹!

img

Rotation์˜ ๊ณต์‹! using Rodrigues Formula

img


์ด๋ฅผ ๋ฒกํ„ฐํ™” ํ•˜๋ฉด

img

img

img

์‹์„ ์ด๋ ‡๊ฒŒ ๋‘๋ฉด d๋ฅผ minimizeํ•˜๋Š” T_k+1์„ ์ฐพ๋Š” ๋ฌธ์ œ๋กœ ๋ฐ”๋€๋‹ค!
๋”ฐ๋ผ์„œ f์˜ ์ž์ฝ”๋น„์•ˆ์„ ๊ตฌํ•œ ๋‹ค์Œ 0์œผ๋กœ ๋งŒ๋“ค์ž

img

LM method๋ฅผ ์ด์šฉํ•ด์„œ Least Square ๋ฌธ์ œ ํ‘ธ๋Š”๊ฒŒ ๋‹ค์ž„

Mapping์€ ์ƒ๋žตํ•œ๋‹ค


Experiment

img

๋ณต๋„๋Š” ๋‹น์—ฐํžˆ ์ž˜ ๋˜๋Š” ๊ฑฐ๊ณ 
์ฑ„์†Œ๋“ค์ด ๋งŽ์€ ๊ณณ๋„ ์€๊ทผ ์ž˜ ๋˜๋Š”๋“ฏโ€ฆ?!


img

img

IMU๋ฅผ ์จ์„œ Kalman Filter ๋Œ๋ฆฐ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋„ ์กด์žฌ~!


Conclusion

LOAM๋…ผ๋ฌธ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ Contribution์ด ์žˆ์Œ

  1. Odometry ์™€ Mapping ๋ชจ๋“ˆ์„ ๋ณ‘๋ ฌ์ ์œผ๋กœ ๋‚˜๋ˆ ์„œ ์ฒ˜๋ฆฌํ–ˆ๋‹ค๋Š” ์ . ๊ฐ 10Hz, 1Hz
  2. Loop closure ์—†์ด low drift & high accuracy๋ฅผ ๋‹ฌ์„ฑํ–ˆ๋‹ค๋Š” ์ 


๋‚˜์˜ ์˜๊ฒฌ

LOAM์€ 2022๋…„ 3์›” ํ˜„์žฌ KITTI Odometry benchmark rand #3์„ ํ•  ์ •๋„๋กœ ์ˆ˜์ค€ ๋†’์€ ์•Œ๊ณ ๋ฆฌ์ฆ˜

ํŠน์ง•์ด๋ผ๊ณ  ๋ฝ‘์•„ ๋†จ์ง€๋งŒ Edge์™€ Planar๊ฐ€ ๋ช…ํ™•ํ•˜๊ฒŒ ๋ถ„๋ฆฌ๋ผ์„œ ์ถ”์ถœ๋˜๋Š”๊ฑด ์•„๋‹˜.. Planar์—๋„ ๋งŽ์€ Edge๊ฐ€ ์ถ”์ถœ๋˜๋”๋ผ


Graph ์ตœ์ ํ™”๊ฐ€ ์•„๋‹Œ ์ด์ƒ LM Method๋กœ ๋งŽ์ด ํ‘ธ๋Š” ๊ฒƒ ๊ฐ™๋‹ค.. ์—ฌ๊ธฐ์„œ ํ•ต์‹ฌ์€ ์–ด๋–ค ๊ฒƒ์„ ๊ธฐ์ค€์œผ๋กœ Data association ํ•˜๋Š๋ƒ์ธ๋“ฏ?


์ด๋Ÿฐ ๋ฅ˜์˜ ๋…ผ๋ฌธ๋“ค์€ ๋ชจ๋‘ ๊นŠ๊ฒŒ ํŒŒ๊ณ ๋“ค๋ฉด ICP์— ๊ธฐ๋ฐ˜ํ•˜๋Š” ๊ฒƒ ๊ฐ™๋‹ค