stem of Pedestrians
r. of Architecture,
ring, pp.195-200. (
itically Tracking
p.257-260., 1996 .
1992. Application
ues to Analysis of
“AIS, June. pp.41 -
I. 1991. Combining
gmentation. 1991
anese )
n Object Extraction
r Space.1992 SNCR
e)
GNZ 35
GNZ 31 GNZ 4:
———— HÀ
GNZ 31—GNZ 43
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 5. Hakodate 1998
AUTOMATED MOSAICING FOR VIDEO IMAGERY CAPTURED FROM MOVING PLATFORMS
Tsuyoshi Kondo, Kazuo Oda,
Masayoshi Obata, Takeshi Doihara
Asia Air Survey Co., Ltd.
8-6 Tamura-cho, Atsugi-shi, KANAGAWA, 243-0016, JAPAN
E-mail : ts.kondo @ari.ajiko.co.jp
kz.oda @ari.ajiko.co.jp
ma.obata @ ari.ajiko.co.jp
ta.doihara @ari.ajiko.co.jp
Commission V, "Working Group IC V / III
KEYWORDS: Automated Mosaicing System, Video Imagery, Moving Platform
ABSTRACT
This study focuses on an automatic image mosaicing system developed in Asia Air Survey Co., Ltd. for video imagery
captured from moving platforms. An automated registration method with non-linear optimization called Levenberg-
Marquardt algorithm is adopted in the system to find transformation coefficients between video frames. The system can
merge series of frames into one image without using additional information such as altitude, location, speed of platforms.
The system calculates translation, rotation and scale factor between frames without specifying corresponding points and
compensates irregularity of motion of a platform. Thus the system can mosaic video imagery which is captured not only
vertically but obliquely to the objects. Mosaicing of Airborne and car-mounted video imagery have been demonstrated by
the system.
1. INTRODUCTION
Automated image mosaicing technique has been studied
for about ten years. Anandan(1987) introduced an
algorithm of automatic image registration with hierarchical
estimation technique, and Hansen et al.(1994)
implemented the algorithm on a hardware based system.
These techniques can merge images into one image
without additional information such as altitude, location,
and speed.
Asia Air Survey Co., Ltd. has developed an automatic
mosaicing system for the video imagery. This system is
designed to run on personal computer and to
automatically register and merge video sequence of up to
10,000 frames at a time. The system gives a quick and
cheap solution to merging series of video frames captured
from moving platforms such as airplanes, helicopters and
cars. The system can process video imagery of vertical
view as well as of oblique view. The output mosaic
images can help us to grasp overall texture of long objects
such as roads, power-transmission lines, and rivers,
which is difficult to visualize in a single video frame.
After short description about the theory of automated
image registration, we will introduce the automated
mosaic system and its demonstration including results for
aerial views of an urban area and power-transmission
lines, and car-mounted camera views of an autobahn
structures and a mall street.
2. THEORY OF 2-D IMAGE REGISTRATION
The algorithm of image registration adopted here is based
on 2-D image mosaicing method which can automatically
match one image with another. This algorithm assumes
that correspondence of coordinates between two images
is registered by projective transformation:
Sy NN
Bix hs
1
C hxthytl
Image 7 | Image I’
Figure1 Corresponding points between the image / and I’
where (x,y) and (x’,y’) are coordinates of Image / and I’,
H(h,..,h) is a set of coefficients of projective
transformation.
Many studies on 2-D image mosaicing adopt Levenberg-
Marquardt (LM) algorithm to calculate H automatically
(Szeliski, 1994). The LM algorithm is a non-linear
optimization which is an extension of least square
minimization (Press et al., 1992). Here LM algorithm
optimizes H which minimize the following evaluation
function:
E=Y {I y)- Ixy) =3¢
where I(x,y) and I'(x’,y’) are pixel value of image / and I".
In many cases 2-D image mosaicing employs coarse-to-
fine strategy which refines precision by processing series
743
