3 1, 1, 1, 1 3D Environment Measurement Using Binocular Stereo and Motion Stereo by Mobile Robot with Omnidirectional Stereo Camera Shinichi GOTO 1, R

Size: px

Start display at page:

Download "3 1, 1, 1, 1 3D Environment Measurement Using Binocular Stereo and Motion Stereo by Mobile Robot with Omnidirectional Stereo Camera Shinichi GOTO 1, R"

しょうこさわまつ
5 years ago
Views:

3 1, 1, 1, 1 3D Envronment Measurement Usng Bnocular Stereo and Moton Stereo by Moble Robot wth Omndrectonal Stereo Camera Shnch GOTO 1, Ryosuke KAWANISHI 1, Atsush YAMASHITA 1 and Toru KANEKO 1 1

accomplsh autonomous tasks. In unknown envronments, they should generate ther envronment maps by themselves.

1 3 1, 1, 1, 1 3D Envronment Measurement Usng Bnocular Stereo and Moton Stereo by Moble Robot wth Omndrectonal Stereo Camera Shnch GOTO 1, Ryosuke KAWANISHI 1, Atsush YAMASHITA 1 and Toru KANEKO 1 1 Department of Mechancal Engneerng, Shzuoka Unversty 3-5-1, Johoku, Naka-ku, Hamamatu-sh, Shzuoka, , Japan Map nformaton s mportant for path plannng and self-localzaton when moble robots accomplsh autonomous tasks. In unknown envronments, they should generate ther envronment maps by themselves. An omndrectonal camera s effectve for envronment measurement, because t has a wde feld of vew. There are bnocular stereo and moton stereo for measurement by omndrectonal camera. However, each method has advantages and dsadvantages. Then, n ths paper, we am to mprove measurement accuracy by usng two omndrectonal cameras nstalled on a moble robot usng bnocular stereo together wth moton stereo. In addton, stereo matchng accuracy s mproved by consderng omndrectonal mage dstorton. Expermental results show the effectveness of our proposed method. Key Words : Omndrectonal Camera, 3D Measurement, Moble Robot 1. 3 (1) ()(3) {f003001,f ,tayamas,tmtkane}@pc.shzuoka.ac.jp (a) Camera (b) Image Fg. 1 Omndrectnal camera (4) 360 1(a) 1(b) (5)(6) 1 (7)

2 Omndrectonal camera Base-lne length :Good result :Bad result Fg. Measurement pont 1 Movement drecton Bnocular stereo Measurement pont Moble robot 1 (8) Structure from Moton SFM 1 (9) 3 (10). 4 Before movement Base-lne length After movement Measurement pont 1 Fg. 3 Moton stereo Measurement pont 5 Image Acquston Measurement Bnocular Stereo Moton Stereo Measurement Data Integraton Fg. 5 Fg. 4 Process Omndrectonal camera Movement drecton Moble robot Camera confguraton r = [x,y,z] T 6 (u,v) (1)() r a b c f

3 s = r = su sv s f c (1) a ( f a + b + b u + v + f ) a f b (u + v ) () Fg. 8 Perspectve mage Hyperbolod Object pont Object X + Y a Z b = 1 r Projecton plane1 c ( = a + b ) Z Projecton plane c Image plane X x Y ( u, v ) f y Projecton center Lens Fg. 6 Calculaton of ray vector 4. Fg. 9 Projectve dstorton 4 1 s y 7 Hyperbolod Hyperbolod Projecton plane d rl Left mage r r s x Rght mage Eppolar lne Fg. 10 Detectng of correspondng pont Camera ray1 Camera ray Eppolar lne Image plane1 Image plane Fg. 7 Eppolar lne 4 (10) d s x s y d 1 r l

4 10 r r r l d r l r r s y d SAD Sum of Absolute Dfference d s x s y Lucas Kanade Tracker (11) (1) 5 3 r = [x,y,z ] T,r = [x,y,z ] T (3) E E (4) (4) (5) r T Er = 0 (4) u T e = 0 (5) u = [x x,y x,z x,x y,y y,z y,x z,y z,z z ] T (6) e = [e 11,e 1,e 13,e 1,e,e 3,e 31,e 3,e 33 ] T (7) e ab E a b E 8 (8) mn e Be (8) B = [u 1,u, u n ] T (9) n E R t = [t x,t y,t z ] T (10) E = RT (10) 0 t z t y T = t z 0 t x (11) t y t x 0 E R t R t 3 3 p Before movement 1 measurement pont (moton stereo) Fg. 11 m After movement p measurement pont Scale matchng (bnocular stereo) 11 p'

7 1 1 360mm 190 1080pxels Fg. 1 Moble robot Fg.

5 p p mp = p (1) m m 6. 1 (7) 3 g (13) g = [g x,,g y,,g z, ] T (13) p g x, = x 1, + p x 1, + p x, + p x, x 1, u 1, x 1, v 1, x, u, x, v, (14) p g x, = y 1, + p y 1, + p y, + p y, y 1, u 1, y 1, v 1, y, u, y, v, (15) p g x, = z 1, + p z 1, + p z, + p z, z 1, u 1, z 1, v 1, z, u, z, v, (16) g (17) h g < h (17) mm pxels Fg. 1 Moble robot Fg. 13 Feature ponts 7 1m KLT SAD (a) Proposed method (b) Prevous method t s n o p g n d n o p s e r o C Fg. 14 Correspondng ponts Tradtonal method Proposed method Fg. 15 Smlarty[%] Hstogram

18 Measurement by bnocular stereo (a) Left camera (b) Rght camera Fg.

6 1 1 1 Fg. 16 Envronment (a) Left mage (b) Rght mage Fg. 17 Stereo mage par (17) mm 578mm 1 3m 4.5m m 0.5m.5m 1.5m Fg. 18 Measurement by bnocular stereo (a) Left camera (b) Rght camera Fg. 19 Measurement by moton stereo Table 1 Processng tmes Measurement Method (Measurement Range) Processng tme [sec] Bnocular Stereo Moton Stereo 6.3 Table Standard devaton from the least square error plane Standard devaton [mm] Plane1 1.1 Plane 59.0

7 3 Fg. 0 Integrated measurement data (No evaluaton value) Plane1 Plane Fg. 1 Integrated measurement data (Top vew) Fg. Integrated measurement data (Brd s eye vew) 8. (1) H. Ishguro, M. Yamamoto and S. Tsuj: Omn- Drectonal Stereo, IEEE Transactons on Pattern Analyss and Machne Intellgence, Vol.14, No., pp.57-6, 199. () C. Geyer and K. Danlds: Omndrectonal Vdeo, The Vsual Computer, Vol.19, No.6, pp , 003. (3) R. Bunschoten and B. Krose: Robust Scene Reconstructon from an Omndrectonal Vson System, IEEE Transactons on Robotcs and Automaton, Vol.19, No., pp , 003. (4) :, D-II, Vol.J84-D-II, No., pp , 001. (5) : MOVIS, 3, 1B15, pp.1 4, 005. (6) :, (MIRU005), pp , 005. (7) R. Kawansh, A. Yamashta and T. Kaneko: Three- Dmensonal Envronment Model Constructon from an Omndrectonal Image Sequence, Journal of Robotcs and Mechatroncs, Vol.1, No.5, pp , 009. (8) :,, Vol.15, No.8, pp , (9) P. Chang and M. Hebert: Omn-drectonal Structure from Moton, Proceedngs of the 000 IEEE Workshop on Omndrectonal Vson, pp , 000. (10) : 3, 10, 1A-E9, pp.1-4, 010. (11) J. Sh and C. Tomas: Good Features to Track, Proceedngs of the 1994 IEEE Computer Socety Conference on Computer Vson and Pattern Recognton, pp , (1) J. Y. Bouguet: Pyramdal Implementaton of the Lucas Kanade Feature Tracker Descrpton of the Algorthm, OpenCV, Intel Corporaton, 000.

Fig Measurement data combination. 2 Fig. 2. Ray vector. Fig (12) 1 2 R 1 r t 1 3 p 1,i i 2 3 Fig.2 R 2 t 2 p 2,i [u, v] T (1)(2) r R 1 R 2

Fig Measurement data combination. 2 Fig. 2. Ray vector. Fig (12) 1 2 R 1 r t 1 3 p 1,i i 2 3 Fig.2 R 2 t 2 p 2,i [u, v] T (1)(2) r R 1 R 2 IP 06 16 / IIS 06 32 3 3-D Environment Modeling from Images Acquired with an Omni-Directional Camera Mounted on a Mobile Robot Atsushi Yamashita, Tomoaki Harada, Ryosuke Kawanishi, Toru Kaneko (Shizuoka