ANALYSIS OF DIFFERENT CALIBRATION MODELS FOR SEE-THROUGH HEAD-
MOUNTED DISPLAYS USED FOR AUGMENTED REALITY APPLICATIONS
Johannes Leebmann
Institute of Photogrammetry and Remote Sensing, University of Karlsruhe (TH), Englerstrasse 7, 76128 Karlsruhe,
Germany - leebmann@ipf.uni-karlsruhe.de
Commission V, WG V/6
KEY WORDS: Augmented Reality, Calibration, Algorithms
ABSTRACT:
There exists a variety of different mathematical models for the calibration of augmented reality systems (ARS). In contrast to camera
calibration, ARS calibration for see-through head-mounted-displays (STHMD) is facing the problem of a low image point accuracy
and other physical circumstances: each image point refers to a image plate with different orientation and position. The orientation
and position is measured for each image point. This different constellation results in a noticeable increase of complexity in the
mathematical model. Up to now the more complex model is circumvented by simplifications in the calibration process. Different
simplified models can be found in literature. On the one hand, the models differ in the way they concatenate the different
parameters. On the other hand, there exist variations in the interpretation of the different parameters as being "observed", "fixed" or
"estimated". The aim of this study is to show how the complexity of the calibration can be handled and to discuss the different
models with different degree of simplification. For the comparison of these models a flexible construction kit is needed, that allows
to describe these differences in the models correctly. Object oriented programming enables the creation of such a construction kit.
The main design concepts of the used construction kit are given. The results show that it not sufficient to assume only the image
points as affected by errors. Far from it, the consideration of the error of the measured orientation of the image ray leads to
significantly different results.
1. INTRODUCTION model. Additionally to this simplification of the stochastic
model discrepancies in the mathematical models can be found:
The goal of an ARS is to superimpose in real time a real world
scenery with a virtual extended version of itself. Such an ARS
is also developed as part of a disaster management tool of the
collaborative research centre 461 (CRC461): "Strong
Earthquakes" [2]. Rescue units are supposed to use the ARS as
a tool to plan their actions on site using the possibilities offered
by virtual reality. In figure 1 a possible scenario is drafted.
Regarding to the reconnaissance strategy of the CRC461 there
will be airborne laser scanning data collected of the whole
affected area after an earthquake. That means that the
geometrical shape of the buildings is known. These three-
dimensional data can be fused with other information available,
e.g. digital elevation model, building structure and so on. This
information can now be used as planning information for rescue
units. The construction of such a system is a challenge in many
ways. One of the problems is the calibration of the used
components. This calibration problem is analysed in detail in
this paper.
An optical see-through augmented reality systems (ARS)
consists in principle of a See- Through Head-Mounted Display
(STHMD) and a head tracking device [5]. The first approaches
of optical see-through AR calibration tried to transfer the
camera calibration procedures known from photogrammetry to
the optical see-through systems. Examples for such approaches
of AR calibration are described in [7] and [11]. These
techniques use simplifications of the real nature of the measured
data. They do not regard that the measurements of the head
tracking device are affected by sensor errors. As a result of this
simplification a large number of observed image points is
necessary (e.g. reported in [11]) to compensate the error in the
the number of used calibration parameters varies from 8 to 11
between the different approaches. The extension of the
mathematical and stochastic model may seem to be only small
modifications. However, a closer look shows that elaborate
modifications of the conventional camera calibration programs
are necessary.
Figure 1. The idea is to use virtual reality for planning rescue
activities. In the figure one can see a virtual cut through a
damaged building.
Only considering the image points as disturbed by errors, one
does not need a-priori knowledge about the real image point
errors. But if a hybrid parameter estimation with different
Observation errors is used (of the image points, tracking
position and angles), the a-priori knowledge essential. The
structure of the mathematical model is more complex as if head
-594—
na
sx Me ^ JJ" N NS "'
i A DMS o
go