'anbul 2004 
1 for recon- 
p. 133-133. 
ant features, 
IEEE Com- 
| many per- 
on - ECCV 
on, Copen- 
t Il, Lecture 
355-369. 
que for self- 
1 Computer 
lo, pp. 112- 
> evaluation 
1 Computer 
E. n.d. 
line stereo 
mputer Vi- 
bratio algo- 
tional Con- 
ystems and 
etric primi- 
ing: Image 
€ Vision to 
n Interface 
ssion. Jour- 
71-880. 
v matching 
1 European 
rk. 
ion. IEEE 
ence 17(8), 
om motion 
" Computer 
notion seg- 
143. 
eo, motion 
BASED ON STEREO SEQUENCE IMAGE 
3-D MOTION PARAMETERS DETERMINATION 
Chunsen ZHANG, Jianging ZHANG, Shaojun He 
School of Remote Sensing and Information Engineering 
Wuhan University, zhchunsen@sina.com 
Commission V, ICWG V/ III 
KEY WORDS:Vision Sciences, Video Sequences, Calibration, Feature Extraction, Stereoscopic Matching, 3-D Motion Parameters 
ABSTRACT: 
Deriving accurate 3-D motion information of a scene is an essential and important task in computer vision, and is also one of the 
most difficult problems. In this paper, using photogrammetry method and computer vision technique the author investigates the 
determination of 3-D motion parameters from binocular stereo image sequences method and steps. Discussing the in-situ calibration 
for binocular stereo computer vision systems, combining correlation coefficients and relaxation method for performing motion and 
binocular stereo matching of images based on point feature, 3-D feature points in the motion object correspondences before and after 
motion, using “Skew-Symmetric Matrix Decomposition (SSMD)” algorithms which exploits the fewest variables and without losing 
the linearity in computation to reduce computing complexity and improve accuracy about motion parameters R and T acquirement 
etc. Finally, the motion parameters of real data based on 3-D correspondences feature estimation method are given. 
1. PREFACE 
It is an important question in computer vision to restore the 
motion information of objects from the sequence image. The 
three-dimensional space coordinates of the feature points of 
moving object is obtained from the binocular stereo sequence 
image. Then, according to the above three dimensional 
positional information(revolve matrix R and translation vector 
T), the moving parameters are figured out. Compared to the 
single sequence image movement analysis method, this 
computation is simple and also can obtain the absolute 
translation quantity. But in this method the spatial three 
dimensional coordinates gain, movement object feature point 
position extraction, effective match between feature point 
stereo and movement(sequence), and 3-D feature points in the 
motion object correspondences before and after motion 
constitute to the key to solve this question. In order to guarantee 
different time movement object feature point correspondence, 
the method presented in this paper at first establish binocular 
stereo vision system, and obtains the translation relationship 
between image space and object space coordinates; then carries 
on sequence and stereo match of binocular stereo sequence 
image of movement object random feature points. The 
calibration result is used to obtaining the object space 
coordinates of feature point, and from this to estimate the 
parameter of object movement (R, T). 
2.BINOCULAR STEREO VISION 
SYSTEM CALIBRATION 
The calibration about camera is an essential procedure in the 
binocular sequence image three-dimensional determination. The 
precision and the reliability of camera calibration directly 
influence 3D measure precision of Stereovision system. In order 
719 
ot fit for such case that the off-the-shelf CCD camera’s interior 
parameters are unstable during the moving object tracking 
process, The author describes the methods of real-time in-suit 
calibration of CCD camera in the process of movement analyses 
based on feature correspondence by use of the known feature 
lines (points) which have geometric relation in object side. This 
calibration model is direct linear transformation (DLT) with 
distortion correction. 
From t Âiek LA AL Y+L,0+L, ES 
Lo + ho) +L, 2+] 
LXWLYTLZYIL, sa 4 
LX wd kl Zl 
Where ( L,,L,,---,L,, Jare 11 unknown DLT parameters, 
(X,Y,Z ) is the space coordinates control point in object 
space. (I, J) is the pixel coordinates, con,” is the pixel 
distortion value, CV;, v ) is the correction values of (I, J) 
when redundancy exists. The quantites of the elements about 
exterior orientation and the photoelectricity transforms 
distortion basic term Nx, Ny coefficient may be obtained from 
Bla Dar 
After L, is obtained, the space coordinates ( U,V,W ) ot 
moving object random feature point’ can be obtained by 
matching result in both left and right images took from stereo 
vision system.Supposes the coefficients L; of two images be 
6 dois 1a 3 and CLi,La,, Lu) respectively. the 
mathematical models for calculating object space feature point 
CUV,W) is: 
Jtv, +8 + 
Lex d El Lt RU Ls 
hh eis bere. trie © 
Lok x4 Latvia be L, +x 
L. + X Ls La zb y Lin La x y Ln PB + y