RELATIVE LASER SCANNER AND IMAGE POSE ESTIMATION FROM POINTS AND SEGMENTS
Matthieu Deveau^ ^, Marc Pierrot-Deseilligny^, Nicolas Paparoditis^, Xin Chen”
“ Institut Géographique National/MATIS Laboratory, 2-4 avenue Pasteur, 94165 Saint-Mandé - firstname.lastname@ign.fr
” MENSI S.A. 30 rue de la Fontaine du Vaisseau, 94120 Fontenay-sous-Bois - firstname.lastname@mensi.fr
KEY WORDS: scanner laser, image data, terrestrial surveying, pose estimation.
ABSTRACT:
This paper presents an approach involving linear features for pose estimation. Here we are interesting in surveys mixing image and
laser scanning, for metrological applications. Since data need to be registered with the best accuracy, we are faced to a 2D-3D pose
estimation problem. In most cases, scenes contain numerous segments, which are good orientation clues. We use these segments to
find pose. Therefore, targets are less prevalent for location and orientation estimation purpose. This means less field operations during
data acquisition. Since some scenes with very few straight lines can leave insufficient spatial constraints, we reintroduce points. We can
deal with feature points to reinforce the system. Then, the algorithm simultaneously minimizes an energy function managing distances
between 3D points projection in images and image points, and distances on segments ends. Precise determination of primitives in 2D
and 3D data leads to fine orientation. Using subpixelar regression after an edge detection gives high-quality estimates for 2D segments.
In point clouds, 3D segments come from plane intersection. We discuss relative influence of features through uncertainty assessment.
1 INTRODUCTION
1.1 Context
Our research study deals with 3D reconstruction of terrestrial
scenes such as cultural heritage main buildings or industrial envi-
ronments. Combining laser data with image is supposed to ease
and automate surface reconstruction. We are betting on geomet-
rical complementarity of such heterogeneous data. Hence, our
goal is to reach a very precise orientation of data sets so as to
perform parallel segmentation of both data. We are presenting
here our work towards image pose estimation relative to a point
cloud involving metric distances between segments and between
points.
For many laser scanner systems nowadays available, a digital
video camera is interdependent with the scanner body. Thus, cal-
ibration is carried out only once (maybe regularly, if needed).
Yet, some of these cameras have got low resolution that gives
poor geometrical information and poor texture for final model.
Furthermore, ratio between image resolution and scan resolution
depends only on scan resolution. With a free camera, one can
choose higher image resolution than scan resolution, without al-
tering scan resolution. Moreover, it allows to take pictures from
different points of view, so to handle occlusions, by means of
convergent photogrammetry. We are also putting ourselves in this
way in a general context. Finally, this frame enables exact model
overlay, with high resolution image.
1.2 Related work
Many solutions have been studied in the pose estimation frame-
work from 2D-3D correspondences. Most ofthe methods coming
from the photogrammetry community use point matches. There
are direct solutions using three, four or six points (Wang and Jep-
son, 1994). More accurate results, when data present noise, come
from least-square resolution with a larger points set. Even in
computer vision, points correspondence remains the most com-
mon pose estimation technique (Haralick et al., 1994). In some
cases, fundamental matrix estimate drives to both intern and ex-
tern parameters compute. Thus, resolution needs more points,
minimum seven, eight for Hartley's algorithm (Hartley, 1997).
One main problem with fundamental matrix estimate comes from
interdependence between intern and extern parameters. If camera
calibration is available, essential matrix estimate provides bet-
ter results. Here, we are considering rigid body digital cameras
which intern parameters are known and computed indepen-
dently. Focal length, principal points and radial distortion are
determined precisely by a calibration procedure on a target field.
1130
As it has been pointed formerly in (Habib, 1999) where one can
find an overview of previous works about pose estimation, man-
made environments are rich in linear features. These features can
be found often on planar intersections. Earlier works (Dhome et
al., 1989) used three lines in the image corresponding to three
ridge lines on the object. (Hanek et al., 1999) have shown that
better results come from exploiting segments ends points rather
than lines. Other approaches (Van den Heuvel, 1999) use geomet-
ric constraints between features. Kumar and Hanson (Kumar and
Hanson, 1994) have studied two main models : one estimates dis-
tances between 3D segment ends projected on image plane and a
line extracted from image ; another minimizes distances between
2D segment ends and the line built on 3D segment ends projec-
tion into image plane. Second model performs better, regarding
to final solution. We have chosen this approach in this field ap-
plication case, in a (photogram)metric context.
1.3 Our approach
[n this framework, segments matching avoids using targets. Till
now, we have been managing surveys with spheres and targets for
points correspondences. To get targets’ center fine position, high
resolution scan is needed, operation which is time-consuming :
much of field work is spent in scanning particular points which
may not be useful in reconstruction...
Moreover, targets can not always be spread correctly all over the
imaged overlaps and thus leads to imprecise geometrical deter-
mination. Thus identifying and matching scene invariants in the
data acquired is a real trend for automating and increasing the
quality of surveys through the quality of pose.
Although using segments reduces time waste, we are sometimes
faced with weak configurations where few segments are present
in the scene. Such scenes can leave indetermination because of
several straight lines in the same plane or in the same direction.
When scanning a facade, many segments lie on the same plane,
and most of them are parallel (vertical or horizontal). For a goth-
ical frontage(our example), they are mostly vertical. In other
cases, there are too few segments to highlight faults. For instance,
with industrial environments made of pipes, perspective avoids
matching on cylinders edges. To overcome these difficulties, we
have chosen to reintroduce points, but without extra field oper-
ation work (sphere or target high resolution scan). We can then
choose feature points where constraints coming from segments
are too weak.
In
My
