art B3. Istanbul 2004
npsac: A synthesis of
census. PAMI 25(3),
| Fitzgibbon, A., 2000.
In: W. Triggs, A. Zis-
sorithms: Theory and
-375.
tomatic relative orien-
n: ISPRS Commission
ometry in Stereo, Mo-
demic Publisher.
Estimating Sensor Pose from Images of a Stereo Rig
a b
Ouided Bentrah" ?, Nicolas Paparoditis^, Marc Pierrot-Deseilligny^, Radu Horaud”
" Institut Géographique National, MATIS, 2-4 avenue Pasteur, 94165 Saint-Mandé. France - firstname.lastname@ign.fr
® INRIA Rhone-Alpes, GRAVIR, 665 avenue de l'Europe, 38330 Monbonnot Saint Martin, France - firstname.lastname(@inrialpes.fr
KEY WORDS: Sensor Pose Estimation, Stereo Rig, Multi-view Imagery, Terrestrial Photogrammetry, Mobile Mapping Systems,
Short Stereoscopic Baseline, Vanishing Points, Dense Stereo Matching.
ABSTRACT:
In this paper we investigate the estimation of pose of images provided by a rigid stereo rig on a mobile mapping system called STERE-
OPOLIS developped at IGN in the ARCHI project of the MATIS laboratory. In this system the terrestrial photographies are a georef-
erencing device. We use the images as a verticality measurement device by finding the vertical and horizontal vanishing points of both
images of the stereo rig at the same time wich improves robustness and accuracy of the estimation. We also use our stereo baseline
as a photogrammetric range measurement provided by dene stereo matching. The relative sensor pose estimation between successive
acquisitions (t) and (t+dt) of the system can thus be seen as the problem of finding the rigid transformation between the two set of
clouds. We achieve this by extracting and matching 3D planes by FFT correlation on corresponding facade orthoimages. The matching
of tie points and segments between images at (t) and (t+dt) is due to the reduction of the search space space given by the surface models
and the relative pose estimated by our 3D plane matching process.
1 USING IMAGES AND STEREOVISION AS A NAVIGATION
SYSTEM
Many works of the Mobile Mapping GIS community tackle the
subject of the production of image data bases with a imaging sys-
tem mounted on a vehicle for georeferenced surveys in urban en-
vironments. The systems presented in the related literature esti-
mate the localisation and the orientation of terrestrial images only
with direct navigation sensors. However, in dense urban areas the
image poses provided by the direct georeferencing devices are al-
tered due to GPS masks and multiple paths. The use of the im-
age as a help for localisation is an issue not tackled in the MMS
community which in general is rather not very familiar with the
photogrammetric and computer vision problems. If real time pro-
cessin is not required, image is a very high quality georeferncing
device.
In this paper we investigate sensor pose estimation of terrestrial
photographies acquired by the mobile Stereopolis system. We
describe an image based georeferencing algorithm for estimating
relative and partial absolute pose which can be derived from a
multi-view rig.
Figure 1: One vertical baselines of the Stereopolis system.
3 THE GEOREFERENCING STRATEGY
2 THE STEREOPOLIS MMS
Es Many researchs of the Mobile Mapping GIS community tackle
The Stereopolis is the MMS system developed at the MATIS lab- lay n ^ ; bping G community £ ke
as : v = A the subject of the production of georeferenced image data bases
oratory of IGN in the ARCHI project for automated acquisition ; 1
= s e eal T uS = with an imaging system mounted on a vehicle for georeferenced
of georeferenced terrestrial images in urban cities. The Stereopo- > 3
surveys in urban environments. Most of the systems presented in
the related literature estimate the localisation and the orientation
of terrestrial images mainly through direct measurements pro-
vided by navigation sensors and mixed with the help of a Kalman
filtering.
lis system as seen in Figure 1, consists of a mobile platform with
three pairs of 4k * 4k CDD cameras and georeferencing devices
(2 GPS with choke rings and an odometer). The cameras are per-
fectly synchronous and dateable (thus allowing a higher quality
of georeferencing at higher speeds). The system provides a good
imaging geometry and good coverage of object space. In our Stereopolis system the image is the output data but is also
the key subsystem for relative and absolute pose estimation as in
computer vision. The GPS in our moving plateform Stereopolis
is only used to reduce the ambiguities of the matches between the
road mark lines reconstructed from the horizontal baseline and
the road mark lines data bases reconstructed from aerial images,
i.e. it provides a very approximate metric measurement which is
: s : used as a gross initial solution. The georeferencing will be pro-
augment the field of view covered by the two cameras. With 28 a M CE AE ; Ze d = p
: M dim. s vided by an “icono-triangulation” by global multi-cameras bun-
mm focal length lenses, the field covers a building six storey high 55 :
En : dle adjustment that integrates measures (tie points and segments)
at a distance of 6 meters (Bentrah et al., 2004). ; il :
from the images and ground control features (GCF) : horizontal
In the scope of the paper we only consider a vertical baseline of and vertical vanishing lines and accurate road lines (zebras, etc.)
this system to estimate the pose as shown in (Figure 1). coming from an external data source basis.
The two frontal facing cameras form a: horizontal stereoscopic
baseline (1.5 meters) allowing the stereo-plotting of urban fea-
tures (lampposts, traffic lights, road marks and signs, trees, etc.
) and two short stereo vertical baselines (1 meter) on each side
of the vehicle to survey the façades on both sides of the street.
The short vertical stereoscopic baselines are slightly divergent to
