i| 2004
erence
raction
tecting
3 map
tecting
; from
ymbine
quality
| with
duced.
raction
je OS
Ikonos
scale in
ited to
| roofs.
instruct
cted by
1986.
is and
visible
mputer
) 2002.
| high
solution
lings of
ig from
)3. CD-
on from
ation to
chnical
theory,
ly 9-12,
;-guided
mmetric
alysis of
y. IEEE
1-326.
surface
with the
mmetric
.344.
REGISTRATION OF AIRBORNE LASER DATA WITH ONE AERIAL IMAGE
Michel ROUX
GET - Télécom-Paris - UMR 5141 LTCI - Département TSI
46 rue Barrault, 75013 Paris - France
michel.roux@enst.fr
KEY WORDS: Airborne Laser Scanner Data, Aerial Image, Image Segmentation, Registration
ABSTRACT
At present, the calibration of airborne laser scanner data relies on the estimation of the position and attitude of the aircraft during the
acquisition using GPS and INS systems, but also on the estimation of some other parameters: time bias, scan angle offset, etc, which
usually requires the acquisition of extra data over known features: along and across the airport runway, over an horizontal building
edge, etc. The operator need then to identify within the cloud of 3D points the position of these known features.
The aim of this paper is to propose a tool for the automated registration of airborne laser scanner data with one aerial image over
urban areas. The method makes use of the intrinsic rigidity of the aerial image: the registration is performed by optimizing the 3D
reconstruction of the scene calculated with the aerial image and the laser points. On the assumption that urban areas are mainly
composed of planar surfaces, a segmentation algorithm generates a partition of the aerial image and a robust technique estimates a 3D
plane for each region. The quality of the registration is calculated according to the global number of outliers remaining after the robust
estimation.
Experimental results show the convexity of this registration estimator for some low frequency deformations: 3D translations and
rotations, and also curvature along and across the flying direction. The system then uses a Nelder-Mead simplex algorithm to calculate
a precise registration of both data sets.
1 INTRODUCTION
The calibration of airborne laser acquisition systems is a complex
task: its goal is first to identify the systematic errors and then to
correct the raw laser data. The different components (the scanner,
the GPS and INS systems, etc.) should be calibrated separately,
as well as the complete mounted system. Part of the calibration
is performed on the ground, but several parameters require also
in flight calibration, which supposes the acquisition of extra data
over perfectly known features before and after, even during, the
acquisition of the data of interest for the mission. The detection
and the estimation of systematic errors from these extra data is
usually carried out interactively, which reduces the automation of
the complete process and causes delays in the delivery of the final
data.
Different studies have been dedicated to the modeling of system-
atic errors in laser acquisition systems (Baltsavias, 1999, Schenk,
2001). Among the identified sources of error, two classes ap-
pear: sources of global deformation of the 3D points cloud, and
sources of local perturbations. For instance, range error generates
a local translation, whose magnitude is constant, but its direction
depends of the relative position of the point and the scanner. On
the other hand, the mis-alignment of the laser system with the
vertical direction generates a global rotation of the 3D points.
The goal of the present paper is to propose an automatic tool to
help for the recovery of systematic errors leading to a global de-
formation of the 3D points. The method is based on the automatic
registration of the laser data with a pre-calibrated aerial image.
The method is based on the estimation of the quality of the 3D
surface reconstruction using both data sets.
A previous study was dealing with the registration of laser data
with a digital elevation model generated from aerial images with
a classical stereo-restitution approach (Bretar et al., 2003). The
new method presented here has the advantage to require only one
1043
image, and to avoid the calculation of a 3D reconstruction from
aerial images which may be the source of additional errors.
The next section presents the image segmentation tool and the
different approaches tested for the estimation of a planar surface
for each region. The section 3 proposes a quality criterion for
the evaluation of a 3D reconstruction calculated with both data.
The convexity of this criterion for different relative deformations
shows its adequation for the automatic registration of one aerial
image with the 3D points acquired by a laser scanner system.
The section 4 presents briefly the approach developed for the au-
tomated registration which is based on a simplex method, and
gives the results of the registration for a scene in the suburb of
Brussels.
2 3D SURFACE ESTIMATION
We first briefly describe the algorithm used for the segmentation
of the aerial image into regions, and then we investigate several
techniques for the estimation of a plane from a set of 3D points.
2.1 Image segmentation
The partitioning of the image into meaningful regions is of great
interest for aerial images of urban areas since it may provide a
useful detection of main components: buildings, roads, cars, etc.,
which will have individual properties in term of geometry and
therefore of models. Region-based segmentation of aerial im-
ages is justified because of the specific properties of urban areas:
man-made structures often have rather constant albedo as they are
often built with a single material.
Split-and-merge techniques for region-based segmentation have
shown their ability to generate consistent and robust partitions.
Among them, the algorithm proposed by Suk and Chung has the
advantage to be a fast and performing technique (Suk and Chung,
1983). This operator is controlled with 3 parameters: each of the
