• ๋Œ€ํ•œ์ „๊ธฐํ•™ํšŒ
Mobile QR Code QR CODE : The Transactions of the Korean Institute of Electrical Engineers
  • COPE
  • kcse
  • ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ๋‹จ์ฒด์ด์—ฐํ•ฉํšŒ
  • ํ•œ๊ตญํ•™์ˆ ์ง€์ธ์šฉ์ƒ‰์ธ
  • Scopus
  • crossref
  • orcid

  1. (School of Electrical and Electronics Engineering, Chung-Ang University, Seoul, Korea)



Camera image, Monocular SLAM, Feature point, Optimization, Kalman filter, RANSAC

1. ์„œ๋ก 

์˜ค๋Š˜๋‚  ์นด๋ฉ”๋ผ ์˜์ƒ์€ ๊ฐœ์ธ์˜ ์ทจ๋ฏธ ํ™œ๋™๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์‚ฐ์—…์˜ ์—ฌ๋Ÿฌ ๋ถ„์•ผ์—์„œ ํ•„์ˆ˜์ ์ธ ์š”์†Œ๋กœ ์ž๋ฆฌ ์žก๊ณ  ์žˆ๋‹ค. ์นด๋ฉ”๋ผ๋Š” ์ปดํ“จํ„ฐ๋น„์ „์˜ ์ฃผ์š” ์„ผ์„œ์ผ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋“œ๋ก , ์ž์œจ์ฃผํ–‰ ์ž๋™์ฐจ์™€ ๊ฐ™์€ ์ด๋™๋กœ๋ด‡์˜ ์„ผ์„œ๋กœ ์ตœ๊ทผ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ๋‹ค. ์ด๋™ ๋กœ๋ด‡์€ ์ž์œจ ์ฃผํ–‰/๋น„ํ–‰์„ ์œ„ํ•ด์„œ ๋กœ๋ด‡์˜ ์œ„์น˜, ์ž์„ธ๋ฅผ ์ถ”์ •ํ•˜๊ณ  ์ฃผ๋ณ€์˜ ์ง€๋„๋ฅผ ๋™์‹œ์— ์ž‘์„ฑํ•˜๋Š” SLAM(simultaneous localization and mapping)์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•œ๋‹ค. ์ตœ๊ทผ ๋กœ๋ด‡ SLAM์— GPS, ๊ด€์„ฑ์„ผ์„œ, ๋ ˆ์ด์ €, ์†Œ๋‚˜ ๋“ฑ ์ „ํ†ต์ ์ธ ์„ผ์„œ ์™ธ์— ์นด๋ฉ”๋ผ๊ฐ€ ์ฃผ์š” ์„ผ์„œ๋กœ ์ž๋ฆฌ์žก๊ณ  ์žˆ๋‹ค.

์ปดํ“จํ„ฐ๋น„์ „ ๋ถ„์•ผ์˜ SfM(structure from motion)๋ฌธ์ œ์™€ ์ด๋™๋กœ๋ด‡์˜ SLAM๋ฌธ์ œ๋Š” ์˜ค๋žœ ๊ธฐ๊ฐ„ ๊ฐ๊ฐ ๋ฐœ์ „ํ•ด ์™”๋Š”๋ฐ, ๋™๊ธฐ๋Š” ๋‹ฌ๋ž์ง€๋งŒ ๊ทผ๋ณธ์ ์œผ๋กœ ๊ฐ™์€ ๋ฌธ์ œ๋ฅผ ๋‹ค๋ฃจ๊ณ  ์žˆ๋‹ค. SfM์€ ์นด๋ฉ”๋ผ์˜ 2์ฐจ์› ์˜์ƒ์œผ๋กœ๋ถ€ํ„ฐ 3์ฐจ์› ์˜์ƒ์„ ์–ป๋Š”๊ฒŒ ๋ชฉ์ ์ด๋ฉฐ ํˆฌ์‚ฌ๊ธฐํ•˜(projective geometry)์™€ ์ตœ์ ํ™” ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜๋Š”๋ฐ ๋น„ํ•ด, SLAM์€ ์—ฌ๋Ÿฌ ์„ผ์„œ๋ฅผ ์‚ฌ์šฉํ•ด์„œ ๋กœ๋ด‡์˜ ์œ„์น˜์™€ ์ฃผ๋ณ€ ์ง€๋„๋ฅผ ์–ป๋Š” ๊ฒƒ์ด ๋ชฉ์ ์ด๊ณ  ํ•„ํ„ฐ๋ง ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•œ๋‹ค. ์ตœ๊ทผ์—๋Š” SfM๊ณผ SLAM์˜ ๋‘ ๋ถ„์•ผ์˜ ๊ฐญ์„ ์ด์–ด์ฃผ๋Š” ์—ฐ๊ตฌ๊ฐ€ ๋งŽ์ด ์ง„ํ–‰๋˜์—ˆ์œผ๋ฉฐ, ์‹ค์‹œ๊ฐ„ SfM๋ฌธ์ œ๋Š” ์˜์ƒ SLAM์ด๋ผ๊ณ ๋„ ํ•œ๋‹ค. SfM์˜ ํŠน์ง•์ ๊ณผ SLAM์˜ ์ง€๋„๋Š” ๊ฐ™์€ ๊ฐœ๋…์ธ๋ฐ, ํ•„ํ„ฐ๋ง๊ธฐ๋ฒ•์—์„œ๋Š” ํŠน์ง•์ ์ด ์ƒํƒœ์‹์˜ ์ƒํƒœ๋ณ€์ˆ˜์— ํฌํ•จ๋˜๋ฏ€๋กœ ํŠน์ง•์ ์ด ๋งŽ์•„์งˆ์ˆ˜๋ก ์ƒํƒœ๋ณ€์ˆ˜์˜ ์ฐจ์ˆ˜๊ฐ€ ์ปค์ง€๋ฉด์„œ ๊ณ„์‚ฐ๋Ÿ‰์ด ๊ธ‰์ฆํ•œ๋‹ค. SLAM์˜ ํ•„ํ„ฐ๋ง ๊ธฐ๋ฒ•๋ณด๋‹ค๋Š” SfM์˜ ์ตœ์ ํ™” ๊ธฐ๋ฒ•์ด ์„ฑ๋Šฅ์ด ์ข‹๋‹ค๋Š” ๊ฒฐ๊ณผ๊ฐ€ ๋ฐœํ‘œ๋˜๊ธฐ๋„ ํ–ˆ๋‹ค(1).

ํŠน์ง•์  ๊ธฐ๋ฐ˜ ์˜์ƒ SLAM์—์„œ๋Š” ์˜์ƒ์—์„œ ํŠน์ง•์ ์„ ์ถ”์ถœํ•˜๊ณ  ๋งค์นญํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ํ•„์š”ํ•˜๋ฉฐ, ์ด์ƒ์ (outlier)์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด RANSAC(RANdom SAmple Consensus)๊ณผ ๊ฐ™์€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•œ๋‹ค. RANSAC์—๋Š” ์ •ํ™•์„ฑ, ๊ณ„์‚ฐ์˜ ์‹ ์†์„ฑ, ๊ฐ•์ธ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์—ฌ๋Ÿฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋“ค์ด ๊ฐœ๋ฐœ๋˜์–ด ์žˆ๋‹ค(2).

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” RANSAC์—์„œ ๋งค์นญ์˜ ์ •ํ™•์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” MSAC(M-estimator)๊ณผ ๋งค์นญ์˜ ๊ณ„์‚ฐ ์†๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” PROSAC (progressive SAC) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฒฐํ•ฉ์‹œํ‚จ MPROSAC์„ ์ œ์•ˆํ•˜์˜€๋‹ค. ์˜์ƒ SLAM์—์„œ์˜ ์นด๋ฉ”๋ผ ์ž์„ธ ์ถ”์ • ๋ฐฉ์‹์—๋Š” ์ „์ฒด์ ์œผ๋กœ 3๊ฐ€์ง€์˜ ๋ฐฉ๋ฒ•(2D-2D, 3D-3D, 3D-2D)์ด ์กด์žฌํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” 2D-2D์™€ 3D-2D์˜ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•œ ์ž์„ธ์ถ”์ • ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๊ณ , ์˜์ƒ SLAM์—์„œ์˜ ์ง€๋„ ์ž‘์„ฑ์„ ์œ„ํ•œ ์ตœ์ ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ ๋ฒˆ๋“ค์กฐ์ •(BA, bundle adjustment)๊ณผ ํ•„ํ„ฐ๋ง๊ธฐ๋ฒ•์—์„œ ๋งŽ์ด ์‚ฌ์šฉ๋˜๋Š” ์นผ๋งŒ ํ•„ํ„ฐ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค.

์ผ๋ฐ˜์ ์œผ๋กœ ๋ถ€์œ ์˜ค์ฐจ(drift)๋ฅผ ์ค„์ด๊ธฐ ์œ„ํ•ด์„œ ๋ฃจํ”„ ํด๋กœ์ง•(loop closing)์„ ํ•˜๊ฒŒ ๋˜๋Š”๋ฐ, ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ฃจํ”„ ํด๋กœ์ง• ๋Œ€์‹ ์— ์ง€์—ญ์ ์ธ ์ง€๋„ ์ž‘์„ฑ๋งŒ ๊ฐ€์ง€๊ณ  ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค.

๋…ผ๋ฌธ ๊ตฌ์„ฑ์€ 2์žฅ์—์„œ ์ขŒํ‘œ๊ณ„ ๋ฐ ์˜์ƒ์˜ ํŠน์ง•์ ์— ๋Œ€ํ•˜์—ฌ ๊ธฐ์ˆ ํ•˜์˜€๊ณ , 3์žฅ์€ ๋ณธ ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•œ MPROSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•˜์—ฌ ๊ธฐ์ˆ ํ•˜์˜€๋‹ค. 4์žฅ์—์„œ๋Š” ์นด๋ฉ”๋ผ ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ•์— ๋Œ€ํ•˜์—ฌ ๊ธฐ์ˆ ํ•˜์˜€๊ณ , 5์žฅ์€ ์นด๋ฉ”๋ผ ์ž์„ธ์™€ 3์ฐจ์› ์ ์— ๋Œ€ํ•œ ๋ณด์ • ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ๊ธฐ์ˆ ํ•˜์˜€๋‹ค. ๋งˆ์ง€๋ง‰ 6์žฅ์—์„œ๋Š” 4์žฅ๊ณผ 5์žฅ์—์„œ ๋‚˜ํƒ€๋‚ธ ๋ฐฉ๋ฒ•๋“ค์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ์‹ค์‹œํ•˜์˜€๋‹ค.

2. ์ขŒํ‘œ๊ณ„ ๋ฐ ์˜์ƒ์˜ ํŠน์ง•์ 

2.1 ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์™€ ์›”๋“œ ์ขŒํ‘œ๊ณ„

์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์™€ ์›”๋“œ ์ขŒํ‘œ๊ณ„์˜ ๊ด€๊ณ„๋Š” ๊ทธ๋ฆผ. 1๊ณผ ๊ฐ™์ด ์›”๋“œ ์ขŒํ‘œ๊ณ„์˜ 3์ฐจ์› ์ ์—์„œ๋ถ€ํ„ฐ ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์˜ ์˜์ƒ ํ‰๋ฉด์˜ 2์ฐจ์› ์ ์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ณผ์ •์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ์œผ๋ฉฐ์‹(1)๊ณผ ๊ฐ™์ด ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 1. ์›”๋“œ์ขŒํ‘œ๊ณ„์™€ ์นด๋ฉ”๋ผ์ขŒํ‘œ๊ณ„์˜ ๊ด€๊ณ„

Fig. 1. World coordinate and camera coordinate

../../Resources/kiee/KIEE.2019.68.1.182/fig1.png

(1)
d u v 1 = f x c u 0 0 f y v 0 0 0 1 r 11 r 12 r 12 t 1 r 21 r 22 r 23 t 2 r 31 r 32 r 33 t 3 X W Y W Z W 1

์—ฌ๊ธฐ์„œ $d$๋Š” depth ๊ฐ’, $u$, $v$๋Š” ์ •๊ทœํ™”๋œ ์˜์ƒ ํ‰๋ฉด์—์„œ์˜ ํŠน์ง•์ , $f _ {x}$, $f _ {y}$๋Š” ์ดˆ์ ๊ฑฐ๋ฆฌ, $c$๋Š” ์™œ๊ณก๊ณ„์ˆ˜, $u _ {0}$, $v _ {0}$๋Š” ๊ด‘ํ•™์ถ•์—์„œ์˜ ์ง์„ ์œผ๋กœ ๋‚ด๋ฆฐ ์˜์ƒ ํ‰๋ฉด์—์„œ์˜ ์ฃผ์ , $X _ {W}$, $Y _ {W}$, $Z _ {W}$๋Š” ์›”๋“œ ์ขŒํ‘œ๊ณ„์—์„œ์˜ 3์ฐจ์› ์ , $r _ { i j } ( i = 1,2,3 j = 1,2,3 )$๋Š” ํšŒ์ „๋ณ€ํ™˜ ํ–‰๋ ฌ์˜ ์„ฑ๋ถ„๋“ค, $t _ { k } ( k = 1,2,3 )$๋Š” ํ‰ํ–‰์ด๋™ ๋ฒกํ„ฐ์˜ ์„ฑ๋ถ„์„ ์˜๋ฏธํ•œ๋‹ค.

2.2 ์˜์ƒ์˜ ํŠน์ง•์  ์ถ”์ถœ, ๊ฒ€์ถœ, ๋ฐ ๋งค์นญ

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ํŠน์ง•์ ์„ ์ถ”์ถœํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ SURF (speeded up robust features)๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค(3). SURF๋Š” ์˜์ƒ์˜ ์ ๋ถ„ ์˜์ƒ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„์‚ฐ ์†๋„๋ฅผ ์ฆ๊ฐ€์‹œํ‚ค๋ฉฐ, ๋ฐ•์Šค ํ•„ํ„ฐ์˜ ์‚ฌ์ด์ฆˆ๋ฅผ ์ฆ๊ฐ€ํ•จ์œผ๋กœ ์ธํ•ด์„œ ์Šค์ผ€์ผ์— ๋ถˆ๋ณ€ํ•˜๊ฒŒ ๋œ๋‹ค. ๋˜ํ•œ ๊ทผ์‚ฌํ™”ํ•œ ํ—ค์‹œ์•ˆ ํ–‰๋ ฌ์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ง•์  ์ถ”์ถœ์„ ํ•˜๊ฒŒ ๋˜๋ฉฐ, ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(2)
H = d x x ( ฯƒ ) d x y ( ฯƒ ) d x y ( ฯƒ ) d y y ( ฯƒ )

(3)
d y x ( ฯƒ ) = โˆ‚ โˆ‚ x โˆ‚ โˆ‚ y ( G ( y , x , ฯƒ ) * f ( y , x ) )

์‹(2)๋Š” ํ—ค์‹œ์•ˆ ํ–‰๋ ฌ $H$, ์‹(3)์€ ๊ฐ€์šฐ์‹œ์•ˆ ๋ผํ”Œ๋ผ์‹œ์•ˆ (Laplacian of Gaussian)์ด๋ฉฐ, $G(x, y, \sigma )$๋Š” ๊ฐ€์šฐ์‹œ์•ˆ ํ•„ํ„ฐ๋กœ ์ž…๋ ฅ ์˜์ƒ $f(y, x)$๊ณผ ์ปจ๋ณผ๋ฃจ์…˜(convolution)ํ•œ ํ›„, y๋ฐฉํ–ฅ์œผ๋กœ ๋ฏธ๋ถ„ํ•˜๊ณ  ๋‹ค์‹œ x๋ฐฉํ–ฅ์œผ๋กœ ๋ฏธ๋ถ„ํ•˜์—ฌ ์–ป์€ ๋„ํ•จ์ˆ˜ $d _{yx} ( \sigma )$๋กœ ํ‘œํ˜„๋œ๋‹ค. ๊ฒ€์ถœ๋œ ํŠน์ง•์ ์ด ํšŒ์ „๋ถˆ๋ณ€ํ•˜๊ธฐ ์œ„ํ•ด ๋ฐฉํ–ฅ(orientation)์„ ์ •์˜ํ•œ๋‹ค. ์ด ๋ฐฉํ–ฅ์€ ํ•˜๋ฅด ์›จ์ด๋ธ”๋ฆฟ ์‘๋‹ต(Haar wavelet response)์— ๋Œ€ํ•œ dx, dy ๋ฐฉํ–ฅ์˜ ์‘๋‹ต์ด ๊ฐ€์žฅ ํฐ ๋ฒกํ„ฐ๋กœ ์ •ํ•ด์ง€๊ฒŒ ๋œ๋‹ค. ์ฃผ ๋ฐฉํ–ฅ์ด ์ •์˜๋˜๋ฉด 4x4 ์œˆ๋„์šฐ๋ฅผ ์”Œ์šฐ๊ณ  ๋‹ค์‹œ ํ•˜๋‚˜์˜ ์œˆ๋„์šฐ์— ๋Œ€ํ•ด 5x5 ์œˆ๋„์šฐ๋ฅผ ์”Œ์›Œ์„œ ๊ฐ๊ฐ์„ dy, dx์œผ๋กœ ํžˆ์Šคํ† ๊ทธ๋žจ์œผ๋กœ ํ‘œ์‹œํ•˜๊ฒŒ ๋œ๋‹ค. ์ด๋ฅผ ์‹์œผ๋กœ ํ‘œํ˜„ํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(4)
V = โˆ‘ d x , โˆ‘ | d x | , โˆ‘ d y , โˆ‘ | d y |

์ด 4๊ฐ€์ง€ ์š”์†Œ๋ฅผ ํžˆ์Šคํ† ๊ทธ๋žจ์œผ๋กœ ๊ทธ๋ ค์„œ ๊ธฐ์ˆ ์ž(descriptor)๋กœ ํ‘œํ˜„ํ•˜๊ฒŒ ๋˜๋ฉฐ, ๊ธฐ์ˆ ์ž๋Š” [$n \times 64$] ์ฐจ์›์˜ ๋ฒกํ„ฐ ํ˜•์‹์œผ๋กœ ์ถ”์ถœ๋œ๋‹ค. ์—ฌ๊ธฐ์„œ n์€ ํŠน์ง•์ ์˜ ๊ฐœ์ˆ˜๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ๋งค์นญ์„ ํ•˜๊ธฐ ์œ„ํ•ด์„œ ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” SSD(sum of squared differences)๋ฐฉ์‹์„ ์‚ฌ์šฉํ•œ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ํ•˜๋‚˜์”ฉ ์ „๋ถ€ ๋‹ค ์กฐ์‚ฌํ•˜๋Š” ์ „์ˆ˜์กฐ์‚ฌ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๊ณ , ์‹์œผ๋กœ ํ‘œํ˜„ํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(5)
S S D ( S u m   o f   S q u a r e s   D i f f e r e n c e ) = D i k - D j k + 1

์—ฌ๊ธฐ์„œ $D$๋Š” ๊ธฐ์ˆ ์ž๋ฒกํ„ฐ(descriptor vector)๋ฅผ ์˜๋ฏธํ•˜๊ณ , $k$๋Š” ์ด๋ฏธ์ง€์˜ ์ˆœ์„œ๋ฅผ ์˜๋ฏธํ•œ๋‹ค. $i$, $j$๋Š” ๊ฐ๊ฐ ์ด๋ฏธ์ง€์˜ ํŠน์ง•์ ์˜ ์ธํ…์Šค๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ๋งค์นญํ•  ๋•Œ SSD์„ ๋Œ€ํ‘œ์ ์œผ๋กœ ์‚ฌ์šฉํ•œ๋‹ค.

๋งค์นญ์„ ํ•˜๊ฒŒ ๋˜๋ฉด ํŠน์ง•์  ๊ฐ„์— ์ด์ƒ์ ๋“ค์ด ๋ฐœ์ƒํ•˜๊ฒŒ ๋œ๋‹ค. ์˜์ƒ ์‚ฌ์ด์˜ ์ด์ƒ์ ์„ ์ œ๊ฑฐํ•ด ์ฃผ๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ๊ฐ€ ์ง€์†๋˜์–ด ์™”์œผ๋ฉฐ(4,5), ๋Œ€ํ‘œ์ ์œผ๋กœ ๋งŽ์ด ์‚ฌ์šฉ๋˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด RANSAC(RANdom SAmple Consensus)์ด๋‹ค. ๋žœ๋คํ•˜๊ฒŒ ์ƒ˜ํ”Œ์„ ๋ฝ‘์•„์„œ inlier์ธ์ง€ outlier์ธ์ง€ ํŒ๋‹จํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋‹ค.

3. MPROSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜

๋ณธ ์ ˆ์—์„œ๋Š” RANSAC์˜ ๊ฐœ์„ ๋œ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋“ค์—์„œ(2) ์†๋„ ์„ฑ๋Šฅ๊ณผ ์ •ํ™•์„ฑ์„ ๋™์‹œ์— ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” MPROSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•ด ๊ธฐ์ˆ ํ•œ๋‹ค.

3.1 ์ œ์•ˆํ•˜๋Š” MSAC/PROSAC ๊ฒฐํ•ฉ ์•Œ๊ณ ๋ฆฌ์ฆ˜

MSAC๊ณผ PROSAC์€ RANSAC์˜ ์„ฑ๋Šฅ์„ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ์„ ๋œ ๋ฐฉ๋ฒ•๋“ค์ด๋ฉฐ, MSAC์€ ์ •ํ™•์„ฑ, PROSAC์€ ๊ณ„์‚ฐ ์†๋„์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ์‹œํ‚จ๋‹ค. MSAC์€ ๊ฒฝ๊ณ„๊ฐ’(threshold)์— ๋Œ€ํ•œ inlier ํŒ์ • ํ›„, outlier๊ฐ€ inlier ์ง‘๋‹จ์— ๋“ค์–ด๊ฐˆ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์—, ๋ชจ๋“  inlier์— ์†ํ•œ ์ง‘๋‹จ์— ๋Œ€ํ•ด ์†์‹ค์„ ๊ณ„์‚ฐํ•˜์—ฌ ์ •ํ™•๋„๋ฅผ ๋†’์ด๊ฒŒ ๋œ๋‹ค. ์†์‹ค(loss) ํ•จ์ˆ˜์—๋Š” Huber ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ์‹(6)์€ Huber ํ•จ์ˆ˜์ด๋‹ค.

(6)
ฯ e 2 = 0 . 5 * e 2 , e 2 < T 2 0 . 5 * T 2 , e 2 โ‰ฅ T 2

์—ฌ๊ธฐ์„œ e๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•œ ์˜ค์ฐจ๊ฐ’, T๋Š” ๊ฒฝ๊ณ„๊ฐ’์ด๋‹ค. PROSAC์€ ์œ ๋„๋œ ์ƒ˜ํ”Œ๋ง ๋ฐฉ๋ฒ•์œผ๋กœ, ์ดˆ๊ธฐ์— ๋งค์นญ์ ์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋งค์นญ ์ ์ˆ˜๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ๊ฒƒ๋ถ€ํ„ฐ ๋‚ด๋ฆผ์ฐจ์ˆœ์œผ๋กœ ์ •๋ ฌํ•˜์—ฌ ๊ฐ€์ค‘์น˜๋ฅผ ์ฃผ๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. ์ด ์ •๋ ฌ๋œ ํ•จ์ˆ˜๋ฅผ quality ํ•จ์ˆ˜๋ผ๊ณ  ๋ถ€๋ฅธ๋‹ค. ๋‹ค์Œ์€ quality ํ•จ์ˆ˜์— ๋Œ€ํ•œ ์‹์ด๋‹ค.

(7)
q u i โ‰ฅ q u j โ‡’ P u i โ‰ฅ P u j ,       i < j

$q( )$์€ quality ํ•จ์ˆ˜, $P$๋Š” ํ™•๋ฅ , $i$, $j$๋Š” ์ž„์˜์˜ ์ˆœ์„œ์ด๋‹ค. ์ด ํ•จ์ˆ˜๋ฅผ ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋งค์นญ๋ฐฉ๋ฒ•์„ ๋‹ค๋ฅด๊ฒŒ ํ•˜์—ฌ ์˜ค์ฐจ ์ˆœ์œผ๋กœ ํ•จ์ˆ˜๋ฅผ ์ •์˜ํ•œ๋‹ค. ๋งค์นญ ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ๊ฐ€์žฅ ์ข‹์€ (first best) ๋งค์นญ์ ๊ณผ ๋‘ ๋ฒˆ์งธ๋กœ ์ข‹์€(second best) ๋งค์นญ์ ์— ๋Œ€ํ•œ ๋น„์œจ๋กœ ๊ณ„์‚ฐํ•˜๋ฉฐ, ๋งค์นญ๋น„์œจ์€ ์‹(8)๊ณผ ๊ฐ™๋‹ค.

(8)
First Best Matching Point Second Best Matching Point = D i k - D j k + 1 D i k - D j ' k + 1

์‹(8)์€ ์‹(5)์— ๋Œ€ํ•ด ๊ฐ€์žฅ ์ข‹์€ ๋งค์นญ์ ๊ณผ ๋‘ ๋ฒˆ์งธ๋กœ ์ข‹์€ ๋งค์นญ์ ์— ๋Œ€ํ•œ ๋น„์œจ์ด๋‹ค. ์ด์™€ ๊ฐ™์ด quality ํ•จ์ˆ˜๋ฅผ ์ ์šฉํ•œ ๋’ค, ์ด์ „์˜ ์ƒ˜ํ”Œ๋ง ์ง‘ํ•ฉ๊ณผ ํ˜„์žฌ์˜ ์ƒ˜ํ”Œ๋ง ์ง‘ํ•ฉ์„ ๋žœ๋คํ•˜๊ฒŒ ๋ฝ‘๋Š”๋‹ค. ์ดํ›„๋กœ๋Š” RANSAC์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ๋™์ผํ•˜๊ฒŒ ์ง„ํ–‰ํ•œ๋‹ค. MPROSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๊ทธ๋ฆผ. 2๊ณผ ๊ฐ™๋‹ค.

๊ทธ๋ฆผ. 2. RANSAC(์ขŒ)๊ณผ MPROSAC(์šฐ) ์•Œ๊ณ ๋ฆฌ์ฆ˜

Fig. 2. RANSAC(left) and MPROSAC(right) algorithm

../../Resources/kiee/KIEE.2019.68.1.182/fig2.png

3.2 MPROSAC ๊ฒฐ๊ณผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜

MPROSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•ด๋ณด๊ธฐ ์œ„ํ•ด์„œ 3.1์ ˆ์— ์†Œ๊ฐœ๋œ MPROSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋งค์นญ ์ „ํ›„์˜ ์‚ฌ์ง„์„ ๊ทธ๋ฆผ. 3์— ๋„์‹œํ•˜์˜€๋‹ค. ์‚ฌ์šฉ ํ›„์— outlier๊ฐ€ ๋งŽ์ด ์ค„์–ด๋‘” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 3. MPROSAC ์‚ฌ์šฉ ์ „(์ขŒ)๊ณผ MPROSAC ์‚ฌ์šฉ ํ›„(์šฐ)

Fig. 3. Before (left) and after (right) using MPROSAC

../../Resources/kiee/KIEE.2019.68.1.182/fig3.png

[ํ‘œ 1]์€ RANSAC, MSAC, PROSAC์„ ๊ฐœ๋ณ„์ ์œผ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ•œ ๊ฒฐ๊ณผ์™€ MPROSAC์˜ ๊ณ„์‚ฐ ์†๋„์™€ ์ •ํ™•๋„ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค. [ํ‘œ 1]์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด MPROSAC ์‚ฌ์šฉ ์‹œ์— ๋งค์นญ์ ์˜ ๊ฐœ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•˜๊ณ , MSAC์— ๋น„ํ•ด์„œ ์˜ค์ฐจ๋Š” ์กฐ๊ธˆ ์ฆ๊ฐ€ํ–ˆ์ง€๋งŒ ๊ณ„์‚ฐ์†๋„๊ฐ€ 30๋ฐฐ ์ด์ƒ ๋นจ๋ผ์กŒ๊ณ , PROSAC์— ๋น„ํ•ด์„œ๋Š” ์˜ค์ฐจ์ •ํ™•๋„์™€ ๊ณ„์‚ฐ ์†๋„ ์„ฑ๋Šฅ์ด ๋ชจ๋‘ ๋‚˜์•„์ง„ ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค.

ํ‘œ 1. RANSAC, MSAC, PROSAC, MPROSAC ์„ฑ๋Šฅ๋น„๊ต

Table 1. Performance comparison among RANSAC, MSAC, PROSAC, and MPROSAC

๋งค์นญ์  ๊ฐœ์ˆ˜

์˜ค์ฐจ[x,y](pix)

์†๋„(sec)

RANSAC

31

[3.34 6.87]

14.12

MSAC

64

[2.47 4.35]

13.21

PROSAC

121

[4.32 7.44]

0.937

MPROSAC

189

[3.12 5.85]

0.432

4. ์นด๋ฉ”๋ผ ์ž์„ธ ์ถ”์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜

4.1 2D-2D ์ž์„ธ ์ถ”์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜

์˜์ƒ ์‚ฌ์ด์— ๋งค์นญ์  ์ถ”์ถœ๊ณผ ์ด์ƒ์  ์ œ๊ฑฐ๋ฅผ ๋งˆ์น˜๋ฉด, ๋‘ ์˜์ƒ ์‚ฌ์ด์˜ ํŠน์ง•์ ๋“ค๋กœ๋ถ€ํ„ฐ ์นด๋ฉ”๋ผ์˜ ์ž์„ธ๋ฅผ ์ถ”์ •ํ•  ์ˆ˜ ์žˆ๋‹ค. ์—ฌ๊ธฐ์— ์‚ฌ์šฉํ•˜๋Š” ๋Œ€ํ‘œ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ 8-point ์•Œ๊ณ ๋ฆฌ์ฆ˜(6)์ด ์žˆ๋‹ค.

๊ทธ๋ฆผ. 4. 2D-2D ์นด๋ฉ”๋ผ ์ž์„ธ ์ถ”์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋ธ”๋ก๋„

Fig. 4. Block diagram for 2D-2D camera pose estimation

../../Resources/kiee/KIEE.2019.68.1.182/fig4.png

๊ทธ๋ฆผ. 4์™€ ๊ฐ™์ด ๋งค์นญ๋œ ํŠน์ง•์ ๋“ค์ด ๋“ค์–ด์˜ค๋ฉด ์—ํ”ผํด๋ผ(epipolar) ์กฐ๊ฑด์‹์„ ์ด์šฉํ•˜์—ฌ Essential ํ–‰๋ ฌ์„ ๊ตฌํ•  ์ˆ˜ ์žˆ๋‹ค. ์—ํ”ผํด๋ผ ์กฐ๊ฑด์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(9)
p ~ ' T E p ~ = 0

์œ„์—์„œ ๋‚˜์˜จ $\overline { p }$, $\overline { p } ^ { \prime }$๋Š” ํŠน์ง•์ , $E$๋Š” Essential ํ–‰๋ ฌ์ด๋‹ค. ์‹(9)์— ๋‘ ์˜์ƒ์—์„œ ๋งค์นญ๋œ 8๊ฐœ์˜ ํŠน์ง•์ ๋“ค์„ ์‚ฌ์šฉํ•˜๋ฉด SVD (singular value decomposition) ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ Essential ํ–‰๋ ฌ์„ ๊ตฌํ•  ์ˆ˜ ์žˆ๋‹ค. ๋˜ Essential ํ–‰๋ น์ด ์ •ํ•ด์ง€๋ฉด ์‹(10)~์‹(13)์œผ๋กœ๋ถ€ํ„ฐ ํ‰ํ–‰์ด๋™ ๋ฒกํ„ฐ $\hat { t }$์™€ ํšŒ์ „๋ณ€ํ™˜ ํ–‰๋ ฌ $R$์„ ๊ตฌํ•  ์ˆ˜ ์žˆ๋‹ค. ํ‰ํ–‰์ด๋™๋ฒกํ„ฐ์ธ $\hat { t }$์€ ๋ฐ˜๋Œ€์นญํ–‰๋ ฌ(skew-symmetric matrix)๋กœ ํ‘œํ˜„๋œ๋‹ค.

(10)
E k = t k ^ R k

(11)
t ^ = U ( ยฑ W ) S U T

(12)
R = U ยฑ W T V T

(13)
W = 0 - 1 0 1 0 0 0 0 1

4.2 ์ƒ๋Œ€์ ์ธ scale ๊ณ„์‚ฐ ๋ฐ ํ‰ํ–‰์ด๋™ ๋ฒกํ„ฐ ๋ณด์ •

4.1์ ˆ์—์„œ ๊ตฌํ•œ ํ‰ํ–‰์ด๋™ ๋ฒกํ„ฐ๋Š” ์ƒ๋Œ€์ ์ธ scale์— ๋Œ€ํ•œ ๋ณด์ •์„ ํ•ด ์ฃผ์–ด์•ผ ํ•˜๋ฉฐ, ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” ์—ฐ์†์ ์ธ 3๊ฐœ์˜ ์˜์ƒ์ด ํ•„์š”ํ•˜๋‹ค. ์ด ๊ณผ์ •์—์„œ 3์ฐจ์› ์ ์„ ๊ตฌํ•˜๊ณ  2๊ฐœ์˜ 3์ฐจ์› ์ ์— ๋Œ€ํ•œ ํŠน์ง•์ ๋“ค์— ๋น„์œจ๋กœ ์ƒ๋Œ€์ ์ธ scale(ํฌ๊ธฐ)๋ฅผ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค. ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ํ‘œํ˜„๋œ๋‹ค.

(14)
r a t i o = X k , ( i , i + 1 ) - X k - 1 , ( i , i + 1 ) X k , ( i - 1 , i ) - X k - 1 , ( i - 1 , i )

$X _{k}$, $X _{k-1}$๋Š” ํŠน์ง•์  k๋ฒˆ์งธ, k-1๋ฒˆ์งธ์˜ 3์ฐจ์› ์ ์„ ์˜๋ฏธํ•œ๋‹ค. $i-1$, $i$, $i+1$์€ ์—ฐ์†์ ์ธ ์˜์ƒ์„ ์˜๋ฏธํ•œ๋‹ค.

4.3 3D-2D ์ž์„ธ ์ถ”์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜

3D-2D ์ž์„ธ ์ถ”์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ง ๊ทธ๋Œ€๋กœ 3์ฐจ์› ์ ๊ณผ 2์ฐจ์› ์ ์œผ๋กœ์˜ ์ž์„ธ๋ฅผ ์ถ”์ •ํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋ฉฐ ๊ทธ๋ฆผ. 5์™€ ๊ฐ™๋‹ค.

๊ทธ๋ฆผ. 5. 3D-2D ์นด๋ฉ”๋ผ ์ž์„ธ ์ถ”์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋ธ”๋ก๋„

Fig. 5. Block diagram for 3D-2D camera pose estimation

../../Resources/kiee/KIEE.2019.68.1.182/fig5.png

์œ„์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋จผ์ € ์—ฐ์†์ ์ธ 3๊ฐœ์˜ ์˜์ƒ์—์„œ 2D-2D ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ์ƒ๋Œ€์ ์ธ ์ž์„ธ๋ฅผ ์ถ”์ •ํ•˜๊ณ  ๊ทธ๊ฒƒ์œผ๋กœ ์ธํ•ด 3์ฐจ์› ์ ์„ ๊ณ„์‚ฐํ•˜๊ณ  ๋‚˜์„œ, 3D-2D ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ตœ์ข…์ ์œผ๋กœ๋Š” 3์ฐจ์› ์ ์„ 2์ฐจ์›์œผ๋กœ ํˆฌ์‚ฌ์‹œ์ผœ์„œ ๋ฐœ์ƒํ•˜๋Š” ์˜ค์ฐจ๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” ์ž์„ธ๋ฅผ ์ฐพ๋Š” ๊ฒƒ์ด 3D-2D ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋‹ค. ์‹์œผ๋กœ ๋‚˜ํƒ€๋‚ด๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(15)
arg min โˆ‘ i p k i - p ^ k - 1 i 2

3์ฐจ์› ์ ์„ ํˆฌ์‚ฌ์‹œํ‚จ ์ ์ด $\hat { p } _ { k - 1 } ^ { i }$ ์œผ๋กœ ํ‘œํ˜„๋˜๊ณ , $p _ { k } ^ { i }$ ์€ k๋ฒˆ์งธ ์˜์ƒ์—์„œ์˜ 2์ฐจ์› ์ ์„ ์˜๋ฏธํ•œ๋‹ค.

4.4 ์‚ผ๊ฐ์ธก๋Ÿ‰๋ฒ•(Triangulation)

3D-2D ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ 2D-2D ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‹คํ–‰ํ•˜๊ณ  ๋‚œ ํ›„, 3์ฐจ์› ์ ์„ ๊ณ„์‚ฐํ•  ๋•Œ ์“ฐ๋Š” ๋ฐฉ๋ฒ•์€ ์‚ผ๊ฐ์ธก๋Ÿ‰๋ฒ•(triangulation)์œผ๋กœ ๋งค์นญ์ ์˜ 3์ฐจ์› ์ขŒํ‘œ๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๋‹ค. ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์—์„œ๋ถ€ํ„ฐ homogeneous ์ขŒํ‘œ๊ณ„(๋™์ฐจ ์ขŒํ‘œ๊ณ„)๋ฅผ ํ‘œํ˜„ํ•˜๋Š” ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(16)
x y 1 = 1 Z X Y Z ,       x ' y ' 1 = 1 Z ' X ' Y ' Z '

x,y๋Š” ์˜์ƒ ํ‰๋ฉด์— ์ƒ๊ธฐ๋Š” ํŠน์ง•์ ์ด๊ณ , X, Y, Z๋Š” ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์—์„œ์˜ 3์ฐจ์› ์ ์„ ์˜๋ฏธํ•œ๋‹ค. ์›”๋“œ์ขŒํ‘œ๊ณ„์˜ 3์ฐจ์› ์ ์„ ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์˜ 3์ฐจ์› ์ ์œผ๋กœ ๋ฐ”๊พธ๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‹์„ ์ •์˜ํ•  ์ˆ˜ ์žˆ๋‹ค.

(17)
P ' = R P + t

$P ^ { \prime }$๋Š” ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์—์„œ์˜ 3์ฐจ์› ์ , $P$๋Š” ์›”๋“œ์ขŒํ‘œ๊ณ„์—์„œ์˜ 3์ฐจ์› ์ , $R$, $t$๋Š” Extrinsic ํ–‰๋ ฌ์˜ ์„ฑ๋ถ„์ธ ํšŒ์ „๋ณ€ํ™˜๊ณผ ํ‰ํ–‰์ด๋™ ๋ฒกํ„ฐ์ด๋‹ค. ๋”ฐ๋ผ์„œ ์ตœ์ข…์ ์œผ๋กœ ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์—์„œ์˜ Z์ ์„ ๊ตฌํ•˜๊ณ  ๊ทธ๊ฒƒ์„ ์›”๋“œ ์ขŒํ‘œ๊ณ„๋กœ ์˜ฎ๊ธฐ๊ธฐ ์œ„ํ•ด ์‹ (17)์„ ์“ฐ๋ฉด ์›”๋“œ ์ขŒํ‘œ๊ณ„์˜ 3์ฐจ์› ์ ์„ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์—์„œ์˜ Z ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(18)
Z 1 ' = r 1 - x r 3 t r 1 - x r 3 x ~ ' ,       Z 2 ' = r 2 - y r 3 t r 2 - y r 3 x ~ '

$Z$์— ๋Œ€ํ•œ ์‹์ด ๋‘ ๊ฐœ ์ด์ƒ์ด ๋‚˜์˜ค๊ฒŒ ๋˜๋ฏ€๋กœ ๋ชจ๋“  ์„ฑ๋ถ„์„ ๋‹ค ๋”ํ•ด์„œ ํ‰๊ท ์„ ๋‚ด๋Š” ํ˜•์‹์œผ๋กœ ์‚ฌ์šฉ๋œ๋‹ค.

5. ์นด๋ฉ”๋ผ ์ž์„ธ ๋ฐ 3์ฐจ์› ์  ๋ณด์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜

5.1 BA(๋ฒˆ๋“ค์กฐ์ •, bundle adjustment) ๋ฐฉ๋ฒ•

BA๋ž€ ๋‹ค๋ฐœ ๋ฌถ์Œ์œผ๋กœ ๋ชจ์•„์„œ ๋ณด์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋ผ๋Š” ๋ง๋กœ ์ง€์—ญ์ (local), ๊ด‘์—ญ์ (global) ๋ฐฉ๋ฒ•์œผ๋กœ ๋‚˜๋‰œ๋‹ค. ๋ณดํ†ต ์˜์ƒ odometry์—์„  ์ง€์—ญ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๊ณ , ์˜์ƒ SLAM์—์„œ๋Š” ๊ด‘์—ญ์ ์œผ๋กœ BA๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค. ์˜์ƒ์˜ ์ž…๋ ฅ์—์„œ๋ถ€ํ„ฐ BA๊นŒ์ง€์˜ ๋ธ”๋ก๋„๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

๊ทธ๋ฆผ. 6. BA ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋ธ”๋ก๋„

Fig. 6. Block diagram for BA algorithm

../../Resources/kiee/KIEE.2019.68.1.182/fig6.png

BA์˜ ๋ชฉ์ ์€ ์นด๋ฉ”๋ผ์˜ ์ž์„ธ์™€ 3์ฐจ์› ์ ์˜ ๋ณด์ •์ด๋‹ค. BA๋Š” ์žฌํˆฌ์‚ฌ ์˜ค์ฐจ๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” ์ž์„ธ์™€ 3์ฐจ์›์˜ ์ ์„ ๊ตฌํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ, ์‹(19)์„ ์ตœ์†Œํ™”ํ•˜๋Š” ์ž์„ธ์™€ 3์ฐจ์› ์ ์„ ๊ตฌํ•œ๋‹ค.

(19)
min โˆ‘ i = 1 n โˆ‘ j = 1 m d Q a j , b i , x i j 2

$a _ {j}$๋Š” ๊ฐ๊ฐ์˜ ์นด๋ฉ”๋ผ ์˜์ƒ $j$๋ฒˆ์งธ๋ฅผ ๋ฒกํ„ฐํ™”, $b _ {i}$๋Š” 3์ฐจ์›์˜ ์ ์˜ ๊ฐœ์ˆ˜ $i$์— ๋Œ€ํ•ด ๋ฒกํ„ฐํ™”, $Q ( )$๋Š” $j$๋ฒˆ์งธ ์˜์ƒ์—์„œ์˜ $i$๋ฒˆ์งธ์˜ ํŠน์ง•์ ์„ ์˜ˆ์ธกํˆฌ์‚ฌํ•œ ํ•จ์ˆ˜๋ฅผ ํ‘œํ˜„, $d ( )$๋Š” ์œ ํด๋ฆฌ๋””์•ˆ ๊ฑฐ๋ฆฌ๋ฅผ ํ‘œํ˜„ํ•œ ๊ฒƒ์ด๋‹ค. $x _ {ij}$๋Š” ์ธก์ •์น˜์˜ 2์ฐจ์› ์ ์„ ์˜๋ฏธํ•œ๋‹ค.

BA์—์„œ์˜ ์ž…๋ ฅ์€ ๋น„์„ ํ˜• ํ•จ์ˆ˜์ด๊ธฐ ๋•Œ๋ฌธ์— ๋น„์„ ํ˜• ์ตœ์†Œ ์ž์Šน์˜ ํ•œ ์ข…๋ฅ˜์ธ Levenberg-Marquardt์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

(20)
J T โˆ‘ X - 1 J + ฮผ I ฮด = J T โˆ‘ X - 1 ฯต

$J$๋Š” ์ž์ฝ”๋น„์•ˆ, $\delta $๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ ๋ฒกํ„ฐ $P$์— ๋Œ€ํ•ด ์—…๋ฐ์ดํŠธ ๋ณ€์ˆ˜, $\mu $๋Š” Dumping ์ธ์ž, $\epsilon = X - \hat { X }$ (์žฌํˆฌ์‚ฌ ์˜ค์ฐจ), ์ด ์˜ค์ฐจ์— ๋Œ€ํ•ด ๋งˆํ• ๋ผ๋…ธ๋น„์Šค ๊ฑฐ๋ฆฌ(Mahalanobis distance) $\epsilon ^ { T } \sum _ { X } ^ { - 1 } \epsilon$์— ๋Œ€ํ•ด LM ๋ฐฉ์‹์˜ ๋น„์„ ํ˜• ์ตœ์†Œ ์ž์Šน ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ํ‘ผ ๊ฒƒ์ด ์œ„์˜ ์‹์ด๋‹ค(7).

5.2 ์นผ๋งŒ ํ•„ํ„ฐ ๋ฐฉ๋ฒ•(Kalman Filter)

๋ณธ ๋…ผ๋ฌธ์€ ํ•„ํ„ฐ๋ง ๋ฐฉ๋ฒ•์œผ๋กœ ํ™•์žฅํ˜• ์นผ๋งŒํ•„ํ„ฐ(EKF, extended Kalman filter)๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์˜์ƒ์ด ์ž…๋ ฅ๋˜๋ฉด ์ดˆ๊ธฐ ํฌ์ฆˆ ๊ฐ’์„ ๋„ฃ์–ด์ฃผ๊ณ , 3์ฐจ์› ์ ์„ ๊ณ„์‚ฐํ•จ์œผ๋กœ์จ ์นด๋ฉ”๋ผ ์ž์„ธ์™€ 3์ฐจ์› ์ ์„ ๋ณด์ •ํ•˜๋Š” ์นผ๋งŒํ•„ํ„ฐ์˜ ์ž…๋ ฅ์ด ๊ฒฐ์ •๋œ๋‹ค. EKF ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๊ทธ๋ฆผ. 7๊ณผ ๊ฐ™๋‹ค.

๊ทธ๋ฆผ. 7. ํ™•์žฅ ์นผ๋งŒํ•„ํ„ฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋ธ”๋ก๋„

Fig. 7. Block diagram for Extended Kalman filter algorithm

../../Resources/kiee/KIEE.2019.68.1.182/fig7.png

EKF์˜ ์ƒํƒœ๋ณ€์ˆ˜ $x$์™€ ์ธก์ •์‹ $I$๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜ํ•œ๋‹ค.

(21)
x = t x t x t y t y t z t z ฮฑ ฮฑ ฮฒ ฮฒ ห™ ฮณ ฮณ

(22)
x k = A x k - 1 + w k ,       w k = N ( 0 , Q )

(23)
I k = h x k + n k ,    n k = N ( 0 , ฮป )

(24)
A = diag 1 ฯ„ 0 1 โ€ฆ , 1 ฯ„ 0 1

(25)
I k = u 1 , v 1 , โ€ฆ u m , v m T

(26)
h x k = f x X 1 Z 1 , f y Y 1 Z 1 , โ€ฆ , f x X m Z m , f y Y m Z m T

์œ„์™€ ๊ฐ™์ด ์ƒํƒœ๋ณ€์ˆ˜๋Š” ์ด๋™ ๋ฒกํ„ฐ์˜ x, y, z ๊ฐ’๊ณผ ์˜ค์ผ๋Ÿฌ ๊ฐ๋„์ธ ๋กค, ํ”ผ์น˜, ์š” ๊ฐ’์„ ์ƒํƒœ๋ณ€์ˆ˜๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ์ƒํƒœ์‹์˜ A ํ–‰๋ ฌ์€ ์นด๋ฉ”๋ผ๊ฐ€ ๋“ฑ์†๋„๋กœ ์ง„ํ–‰ํ•œ๋‹ค๊ณ  ๊ฐ€์ • ํ•˜์— ๊ตฌํ–ˆ๋‹ค. h() ํ•จ์ˆ˜๋Š” ์˜์‚ฐ ํ‰๋ฉด์ƒ ํ‘œํ˜„๋˜๋Š” ์ดˆ์ ๊ฑฐ๋ฆฌ์™€ ์นด๋ฉ”๋ผ ์ขŒํ‘œ๊ณ„์˜ 3์ฐจ์› ์ ์˜ ๊ด€๊ณ„๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋‹ค(8).

6. ์นด๋ฉ”๋ผ ์ž์„ธ์™€ 3์ฐจ์› ์ ์— ๋Œ€ํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜

6.1 ์นด๋ฉ”๋ผ ์ž์„ธ์ถ”์ • ๋ฐฉ๋ฒ• ๋น„๊ต

6.1.1 2D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ

2D-2D ์ž์„ธ์ถ”์ •์˜ ์˜์ƒ์œผ๋กœ๋Š” ๋“ฑ์†๋„๋กœ ์ด๋™ํ•˜๋Š” ๋“œ๋ก ์— ๋Œ€ํ•œ ์˜์ƒ[720x1280]์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๊ทธ๋ฆผ. 8์—์„œ ๋ณด๋Š” ๊ฒƒ๊ณผ ๊ฐ™์ด 2D-2D ์ž์„ธ ์ถ”์ •์˜ scale ๋ณด์ •์„ ํ•˜์ง€ ์•Š์œผ๋ฉด ์˜ค์ฐจ๊ฐ€ x์ถ• ์ง„ํ–‰๋ฐฉํ–ฅ์œผ๋กœ scale์ด ์ปค์ง€๋Š” ๊ฒƒ์„ ํ™•์ธํ•˜์˜€๋‹ค.

๊ทธ๋ฆผ. 8. 2D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ• ๊ฒฐ๊ณผ

Fig. 8. Result of 2D-2D camera pose estimation

../../Resources/kiee/KIEE.2019.68.1.182/fig8.png

6.1.2 2D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ• scale ๋ณด์ • ํ›„ ์„ฑ๋Šฅ

6.1.1์ ˆ์—์„œ ์–ธ๊ธ‰ํ•œ ๋ฐ”์™€ ๊ฐ™์ด 2D-2D ์ž์„ธ ์ถ”์ •์—์„œ scale ๋ณด์ •์„ ํ•˜์ง€ ์•Š์œผ๋ฉด ์ง„ํ–‰ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ์˜ ๊ฑฐ๋ฆฌ๊ฐ€ ๋ฉ€์–ด์ง€๊ฒŒ ๋œ๋‹ค. 4.2์ ˆ์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ๋Œ€์ ์ธ scale ๋ณด์ •์„ ํ•ด์ฃผ์–ด ์–ป์€ ๊ฒฐ๊ณผ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. ๊ทธ๋ฆผ8์— ๋น„ํ•ด์„œ ๊ทธ๋ฆผ. 9๋Š” ์นด๋ฉ”๋ผ์˜ ์œ„์น˜๊ฐ€ ์ž˜ ๋ณด์ •๋˜์—ˆ๊ณ  ์ง„ํ–‰ ๋ฐฉํ–ฅ์ธ x์ถ• ์˜ค์ฐจ๊ฐ€ ๋งŽ์ด ์ค„์–ด๋“  ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 9. Scale ๋ณด์ • ํ›„ 2D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ• ๊ฒฐ๊ณผ

Fig. 9. Result of 2D-2D camera pose estimation after scale correction

../../Resources/kiee/KIEE.2019.68.1.182/fig9.png

6.1.3 3D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ•์˜ ์„ฑ๋Šฅ

์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ• ์ค‘์—์„œ ์ •ํ™•๋„๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ๊ฒฐ๊ณผ๋Š” 3D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ•์ด๋‹ค. 6.1.1์ ˆ์˜ 2D-2D ์ž์„ธ์ถ”์ • ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ x์ถ• ์˜ค์ฐจ๋ฅผ ๋ณด๋ฉด, ๊ทธ๋ฆผ. 10์˜ ๊ฒฐ๊ณผ์—์„œ ์ •ํ™•๋„๊ฐ€ ์ข‹์•„์ง„ ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 10. 3D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ• ๊ฒฐ๊ณผ

Fig. 10. Result of 3D-2D camera pose estimation

../../Resources/kiee/KIEE.2019.68.1.182/fig10.png

6.2 ์นด๋ฉ”๋ผ ์ž์„ธ ๋ฐ 3์ฐจ์› ๋ณด์ • ์„ฑ๋Šฅ

6.2.1 3D-2D๊ธฐ๋ฐ˜ ๋ฒˆ๋“ค์กฐ์ • 3์ฐจ์› ์  ์„ฑ๋Šฅ

6.1์ ˆ์˜ ์ž์„ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ์ •ํ™•๋„๊ฐ€ ๊ฐ€์žฅ ๋†’์€ 3D-2D ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์นด๋ฉ”๋ผ ์ž์„ธ์™€ 3์ฐจ์› ์ ์— ๋Œ€ํ•˜์—ฌ BA ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‹คํ–‰ํ•˜์˜€๋‹ค. ๊ทธ๋ฆผ. 11์„ ๋ณด๋ฉด ์šฐ์ธก์˜ ์‹ค์ œ ์˜์ƒ์˜ ๋ฌผ์ฒด์˜ ์œค๊ณฝ์ด ์ขŒ์ธก์— 3์ฐจ์› ํŠน์ง•์ ์œผ๋กœ ์ž˜ ๋“œ๋Ÿฌ๋‚˜๊ณ  ์žˆ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 11. 3D-2D ๊ธฐ๋ฐ˜ ๋ฒˆ๋“ค์กฐ์ • ํ›„ 3์ฐจ์› ์  ์„ฑ๋Šฅ ๊ฒฐ๊ณผ

Fig. 11. Result of 3D points after 3D-2D based BA

../../Resources/kiee/KIEE.2019.68.1.182/fig11.png

6.2.2 ํ™•์žฅ ์นผ๋งŒํ•„ํ„ฐ 3์ฐจ์› ์  ์„ฑ๋Šฅ

5.2์ ˆ์—์„œ ์–ธ๊ธ‰ํ•œ EKF ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ 3์ฐจ์› ์ ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ์–ป์€ ๊ฒฐ๊ณผ๋ฅผ ๊ทธ๋ฆผ. 12์— ๋„์‹œํ•˜์˜€๋‹ค. ์ขŒ์ธก ๊ทธ๋ฆผ์—์„œ 3์ฐจ์› ์ ๋“ค์ด ํฌ๋ฏธํ•˜๊ฒŒ ๋ณด์ด๋Š”๋ฐ ์˜ค๋ฅธ์ชฝ ๋ฌผ์ฒด์˜ ์œค๊ณฝ์„ ์•Œ์•„๋ณด๊ธฐ ์–ด๋ ต๋‹ค. ์œก์•ˆ์ƒ์œผ๋กœ๋„ ๊ทธ๋ฆผ. 11๊ณผ ๊ทธ๋ฆผ. 12์—์„œ EKF๊ฐ€ BA๋ณด๋‹ค 3์ฐจ์› ์ ์„ ๋ณด์ •ํ•˜๋Š” ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง„๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ๋‹ค.

๊ทธ๋ฆผ. 12. ํ™•์žฅ ์นผ๋งŒํ•„ํ„ฐ 3์ฐจ์› ์  ์„ฑ๋Šฅ ๊ฒฐ๊ณผ

Fig. 12. Result of 3D points after EKF

../../Resources/kiee/KIEE.2019.68.1.182/fig12.png

7. ๊ฒฐ ๋ก 

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋‘ ์˜์ƒ์˜ ๋งค์นญ ํ›„ ์ด์ƒ์ ์„ ์ œ๊ฑฐํ•˜๋Š” RANSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ค‘์—์„œ MSAC๊ณผ PROSAC์˜ ์žฅ์ ์„ ๊ฒฐํ•ฉํ•œ MPROSAC ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ œ์•ˆํ•˜๊ณ  ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•จ์„ ํ™•์ธํ•˜์˜€๋‹ค.

๋˜ ์˜์ƒ SLAM์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ํ•„ํ„ฐ๋ง ๋ฐฉ๋ฒ•์ธ EKF์™€ ์ตœ์ ํ™” ๋ฐฉ๋ฒ•์ธ BA์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค. ํฌ๊ฒŒ 2๊ฐ€์ง€์˜ ์„ฑ๋Šฅ์„ ๋น„๊ต, ๋ถ„์„ํ•˜์˜€๋Š”๋ฐ, ์ฒซ์งธ๋กœ๋Š” ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์„ฑ๋Šฅ ๋น„๊ต, ๋‘˜์งธ๋กœ๋Š” ์นด๋ฉ”๋ผ ์ž์„ธ ์ถ”์ • ๋ฐ 3์ฐจ์› ์  ๋ณด์ •๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค. ์ฒซ ๋ฒˆ์งธ์˜ ๊ฒฐ๊ณผ์—์„œ๋Š” 2D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ•๋ณด๋‹ค๋Š” 3D-2D ์ž์„ธ ์ถ”์ •๋ฐฉ๋ฒ•์ด ์ •ํ™•๋„๊ฐ€ ๋†’๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ๊ณ , ๋‘ ๋ฒˆ์งธ์˜ ๊ฒฐ๊ณผ์—์„œ๋Š” ์นผ๋งŒํ•„ํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ 3์ฐจ์› ์  ๋ณด์ • ๋ฐฉ๋ฒ•๋ณด๋‹ค๋Š” ๋ฒˆ๋“ค ์กฐ์ •์˜ ๋ฐฉ๋ฒ•์œผ๋กœ 3์ฐจ์› ์ ์„ ๋ณด์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ํŠน์ง•์ ์˜ ์ •๋ณด์™€ ์ •ํ™•๋„ ๋ฉด์—์„œ ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ๋‹ค.

๊ฐ์‚ฌ์˜ ๊ธ€

๋ณธ ๋…ผ๋ฌธ์€ 2018๋…„ ์ •๋ถ€(๊ต์œก๋ถ€)์˜ ์žฌ์›์œผ๋กœ ํ•œ๊ตญ์—ฐ๊ตฌ์žฌ๋‹จ์˜ ์ง€์›์„ ๋ฐ›์•„ ์ˆ˜ํ–‰๋œ ๊ธฐ์ดˆ์—ฐ๊ตฌ์‚ฌ์—…(No. 2017R 1D 1A 1B03035207).

References

1 
Strasdat H., Montiel J. M. M., Davison A. J., 2012, Visual SLAM: Why filter?, Image and Vision Computing, Vol. 30, pp. 65-77DOI
2 
Choi Sunglok, Kim Taemin, Yu Wonpil, September 2009, Performance evaluation of RANSAC family, British Machine Vision Conference, BMVC 2009, London, UK,, pp. 81.1-81.12, 7-10Google Search
3 
Bay Herbert, Tuytelaars Tinne, Gool Luc Van, June 2008, SURF: Speeded Up Robust Features, Computer Vision and Image Understanding, Vol. 110, No. 3, pp. 346-359DOI
4 
Fischler Martin A., Bolles Robert C., June 1981, Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Carto-graphy, Communications of the Association for Computing Machinery, Vol. 24, No. 6, pp. 381-395DOI
5 
Torr P. H. S., Zisserman A., 2000, MLESAC: A new robust estimator with application to estimating image geometry, Computer Vision and Image Understanding, Vol. 78, pp. 138-156DOI
6 
Hartley Richard I., June 1997, In Defense of the Eight-Point Algorithm, IEEE Transactions on Pattern Recognition and Machine Intelligence, Vol. 19, No. 6, pp. 580-593Google Search
7 
LOURAKIS MANOLIS I. A., ARGYROS ANTONIS A., March 2009, SBA: A Software Package for Generic Sparse Bundle Adjustment, ACM Transaction on Mathematical Software(TOMS), Article 2, Vol. 36, No. 1, pp. 2:1-2:30DOI
8 
Ragab M. E., Wong K. H., May, 2007, Extended Kalman Filter Based Pose Estimation Using Multiple Cameras, The CSE Dept., The Chinese University of Hong Kong, Internal report, Vol. 16Google Search
9 
Jeon Jin-Seok, Kim Hyo-Joong, Shim Duk-Sun, Oct, 2018, MSAC/ PROSAC Fusion Algorithm to Enhance SURF Per- formance, Conference on Information and Control Systems, Vol. 26, pp. 276-277Google Search

์ €์ž์†Œ๊ฐœ

์ „ ์ง„ ์„ (Jin-Seok Jeon)
../../Resources/kiee/KIEE.2019.68.1.182/au1.png

2017๋…„ ๋‹จ๊ตญ๋Œ€ํ•™๊ต ์ „์ž์ „๊ธฐ๊ณตํ•™๋ถ€ ์กธ์—…(๊ณตํ•™์‚ฌ)

ํ˜„์žฌ ์ค‘์•™๋Œ€ํ•™๊ต ์ „์ž์ „๊ธฐ๊ณตํ•™๋ถ€ ์„์‚ฌ๊ณผ์ • ์žฌํ•™์ค‘

๊ด€์‹ฌ๋ถ„์•ผ๋Š” ๋กœ๋ด‡ SLAM, ๋กœ๋ด‡๋น„์ „, ์œ„์„ฑํ•ญ๋ฒ•

๊น€ ํšจ ์ค‘ (Hyo-Joong Kim)
../../Resources/kiee/KIEE.2019.68.1.182/au2.png

2018๋…„ ์ค‘์•™๋Œ€ํ•™๊ต ์ „์ž์ „๊ธฐ๊ณตํ•™๋ถ€ ์กธ์—…(๊ณตํ•™์‚ฌ)

ํ˜„์žฌ ์ค‘์•™๋Œ€ํ•™๊ต ์ „์ž์ „๊ธฐ๊ณตํ•™๋ถ€ ์„์‚ฌ๊ณผ์ • ์žฌํ•™ ์ค‘

๊ด€์‹ฌ๋ถ„์•ผ๋Š” ๋กœ๋ด‡ SLAM, ์˜์ƒ-๊ด€์„ฑ odometry, ๋กœ๋ด‡๋น„์ „

์‹ฌ ๋• ์„  (Duk-Sun Shim)
../../Resources/kiee/KIEE.2019.68.1.182/au3.png

1984๋…„, 1986๋…„ ์„œ์šธ๋Œ€ ์ œ์–ด๊ณ„์ธก๊ณตํ•™๊ณผ ์กธ์—…(๊ณตํ•™์‚ฌ, ๊ณตํ•™์„์‚ฌ)

1993๋…„ ๋ฏธ์‹œ๊ฐ„๋Œ€ ํ•ญ๊ณต์šฐ์ฃผ๊ณตํ•™๊ณผ ์กธ์—…(๊ณตํ•™๋ฐ•์‚ฌ)

1994๋…„ ๋ฏธ์‹œ๊ฐ„๋Œ€ ์ „๊ธฐ๋ฐ์ปดํ“จํ„ฐ๊ณตํ•™๊ณผ ํฌ์ŠคํŠธ๋‹ฅ

1995๋…„ 3์›”~ํ˜„์žฌ ์ค‘์•™๋Œ€ํ•™๊ต ์ „์ž์ „๊ธฐ๊ณตํ•™๋ถ€ ๊ต์ˆ˜

2009๋…„~2011๋…„ JEET ์—๋””ํ„ฐ

2014๋…„~2017๋…„ IJCAS ์—๋””ํ„ฐ

2018๋…„ ๋Œ€ํ•œ์ „๊ธฐํ•™ํšŒ ๋ถ€ํšŒ์žฅ, ์ •๋ณด ๋ฐ ์ œ์–ด ๋ถ€๋ฌธํšŒ์žฅ

๊ด€์‹ฌ๋ถ„์•ผ๋Š” ์ œ์–ด, ์œ„์„ฑํ•ญ๋ฒ•, ๊ด€์„ฑํ•ญ๋ฒ•, ๋กœ๋ด‡ SLAM, ์˜์ƒ-๊ด€์„ฑ odometry, ๋กœ๋ด‡๋น„์ „