rage
[etric
nent.
'ased
mage
ction
tions
f this
t. By
erent
ethod
imax
sfully
is the
duces
One
n car
or the
ndard
Lm
a
(D
0
Wn
157
RELATIVE ORIENTATION OF TWO DISPARITY MAPS IN STEREO VISION
Olli Jokinen and Henrik Haggrén
Helsinki University of Technology, Institute of Photogrammetry and Remote Sensing
Otakaari 1, FIN-02150 Espoo, Finland
Phone: +358-0-4513907, Fax: +358-0-465077, E-mail: ojokinen@leka.hut.fi
KEYWORDS: Stereo vision, disparity map, relative orientation, modeled features, surface matching.
ABSTRACT
Two methods to solve the relative orientation of two disparity maps measured by a movable stereo head from
distinct viewpoints are presented. The first method is based on modeled features such as plane normals, axes
of cones and cylinders, and vertices of cones. In the second method, the first disparity map is projected onto
the second one and the difference between the projected and the second map is minimized. The methods are
tested with synthetic data.
1. INTRODUCTION
Stereo vision is concerned with recovering depth from two images taken from different viewpoints. The depth
information is obtained as disparity between corresponding points in the rectified left and right image. Deter-
mining disparity for every pixel in the rectified left image results in a disparity map which contains information
about the object space visible to the current location of the camera pair, i.e., the stereo head. When considering
an industrial close-range application where one wants to measure an object from all directions, either the stereo
head or object should be moved. For every viewpoint, a new disparity map is thus obtained. In this paper, we
consider the case of two disparity maps and solve the relative orientation of these maps. The interior orientation
of the stereo head (relative orientation of the cameras) is assumed to be known or determined in advance so that
we solve only the relationship between the exterior orientations of the stereo head in two different positions.
The traditional way of solving exterior orientation of a stereo head is to measure every time the stereo head is
moved particular control points fixed to an external laboratory coordinate system. In some cases, however, there
is no external frame available and all we have in a disparity map is information about the relative position of the
object with respect to the stereo head. Consequently, the relative orientation of two disparity maps should be
solved using only information available in the object itself. In this paper, we present two alternative methods to
do this. In the first method, planar and quadratic regions in the both disparity maps are modeled and modeled
features are then used to estimate the orientation. In the second method, all the original measurements are
matched in the overlap region to give the best fitting orientation. These two methods are presented in Sections
3 and 4, respectively, and Section 5 contains test results with synthetic data. We first review, however, some
useful formulas related to the formulation of the problem. :
2. USEFUL FORMULAS
Consider first a single disparity map measured by a stereo vision system. In industrial close-range applications,
such a vision system could consist of two CCD video cameras connected to a photogrammetric station and
supported by a feature projector (Haggrén et al, 1993). During the measuring process, the left and right image
are rectified to the normal case of stereography. The right-handed X,Y, Z coordinate system is fixed to the
fictitious left camera in the normal case so that the origin is in the projection center, the negative Z-axis points
into the direction of sight, and the X-axis is parallel to the stereo baseline. The left image coordinates z,y and
disparity p are given by
HX Hy HB
E zs uem y ; (1)
where H is the distance from the projection center to the rectification plane and B is the base, i.e., the distance
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995