USING 3-D VIRTUAL REALITY MODELS FOR IMAGE ORIENTATION
IN MOBILE COMPUTING
Charalampos Georgiadis, Anthony Stefanidis, Peggy Agouris
Dept. of Spatial Information Science and Engineering
348 Boardman Hall
University of Maine, USA
{harris, tony, peggy } @spatial.maine.edu
Commission V, WG V/5
KEY WORDS: Close Range, Motion Imagery, Orientation, Virtual Models, Modelling
ABSTRACT:
Advances in sensor technology and computing capabilities and modalities are revolutionizing close-range image collection and
analysis for geospatial applications. These advances create the need for new ways of handling and processing video datasets at quasi
real time rates. In this paper we present an innovative two step orientation technique for ground level motion imagery using a 3
dimensional virtual model as control information. In the first step few select anchor frames are orientated precisely via an image
orientation-through-queries approach. In the second step, intermediate frames are orientated relatively to these anchor frames
through an innovative analysis of building facade variations in them. Combined, these two steps comprise a complete approach to
motion imagery orientation using a VR as control information.
1. INTRODUCTION
During the last years we are experiencing great advances in
sensor technology and wireless communications. That progress
created new data collection schemes, in which users can collect
data roaming a scene with a GPS enabled digital camera or
camcorder. These collection schemes create a vast amount of
data that has to be stored and processed. As a result new
techniques have to be developed that allow for fast processing
of motion imagery, either offline or in a quasi real time manner.
In the context of this paper we use the term motion imagery to
refer to imagery collected at video rates, or even as select
frames captured a few seconds (or even minutes) apart, using
either a video or a still camera. Examples of such datasets
include imagery collected by hand-held cameras captured while
roaming an urban environment, or imagery collected by a
network of fixed sensors (e.g. surveillance cameras) monitoring
a scene. The computing capabilities of mobile devices are also
improving with the advances in technology, creating the
opportunity of location based computing. In order to be able to
acquire modeling information from such data we have to use
some kind of control information, the traditional techniques of
acquiring control points are still time consuming. In addition
research advances in the development of 3 dimensional virtual
models of large scale complex urban scenes have resulted in the
creation of impressive and complex VR models. These
advances provide opportunities for the integration of such
models in motion imagery analysis. The accuracy and
complexity of these models provide an excellent use as control
information for motion image processing.
In this paper we present an approach for the recovery of sensor
orientation and position information using a two-step
procedure. We focus on the use of motion imagery datasets (at
quasi-video rates) captured by sensors roaming an urban scene.
In such a scenario the orientation variations among successive
frames are small. The first step of our approach entails the
comparison of object configurations depicted in a frame to
corresponding configurations identified in the VR model. This
provides excellent approximate values that are then refined
(using precise matching) to provide the orientation parameters
for that frame. This is a process that can be performed for few
anchor frames (e.g. every few minutes) to provide accurate
orientation information at these instances. For subsequent
frames we estimate their orientation by determining their
variation from the nearest anchor frames. To do so we use 2-D
transformations of objects depicted in these frames, and
orientation differences between the frames computed using
vanishing points to translate these transformation parameters to
orientation variations. Combined, these two steps comprise a
novel process of progressive orientation recovery that meets the
computational requirements of mobile mapping applications.
Most of the known techniques for computing relative
orientation need points in different planes. In [Simon and
Berger, 2002] a similar approach is presented in which they
estimate the orientation of an image using the previous image of
known orientation and a planar homography between these two
images. In another approach [Chia et al, 2002], compute the
relative orientation based on one or two reference frames,
exploiting epipolar geometry and using recursive methods. Our
approach tries to solve this problem using points from only one
plane, and using linear methods. Furthermore the information
acquired during the orientation process can be used to update
the existing virtual model. Such update procedure includes
change detection both in geometric and radiometric content, in
the existing objects and detection of new objects or deletion of
objects in the model. In our paper we present the approach for
the orientation estimation in the intermediate frames.
The paper is organized as follows. In section 2 we present an
overview of our navigation-through-virtual models approach. In
Section 3 we present the indices we use for comparing the
intermediate frames with anchor frames for orientation
recovery. Experimental results in section 4 demonstrate the
performance of our approach and conclude with future work
plans in section 5.
Internation
UE Te
2. MC
We assum
scene that
database. S
GPS sensc
position in
pose and uj
Our approa
-Uthe:!
to detern
select an
— the
remainin
Thus we pr
parameters
corrections
orientation
|. Anchor
intervals (i
determined
Anchor Fri
As we can
MPEG coi
frames in
frames, anc
the frames
between c
philosophy
position in
anchor fra
recovered b
size of obje
In figure 2
approach.
corresponds
intermediat
frame orier
presented ir
we should 1
image quer
Classic ima
database ba
WC use ima
