151
In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). 1APRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010
EXTRACTION OF ACCURATE TIE POINTS
FOR AUTOMATED POSE ESTIMATION OF CLOSE-RANGE BLOCKS
L. Barazzetti*, F. Remondino**, M. Scaioni*
* Politecnico di Milano, Department of Building Engineering Science and Technology, Milan, Italy
Email: (luigi.barazzetti, marco.scaioni)@polimi.it, web: http://www.best.polimi.it
** 3D Optical Metrology Unit, Bruno Kessler Foundation, Trento, Italy
Email: remondino@fbk.eu, web: http://3dom.fbk.eu
Commission III - WG 1
KEY WORDS: Feature Extraction, Matching, Orientation, Robust Estimation
ABSTRACT:
The article presents a powerful and automated methodology to extract accurate image correspondences from different kinds of close-
range image blocks for their successive orientation with a bundle adjustment. The actual absence of a commercial solution able to
automatically orient markerless image blocks confirms the still open research in this field. The developed procedure combines
different image processing algorithms and robust estimation methods in order to obtain accurate locations and a uniform distribution
of tie points in the images. We demonstrate the capabilities and effectiveness of this method with several tests on closed or open
sequences and sparse blocks of images captured by a standard frame (pinhole) camera, but also spherical images. An accuracy
evaluation of the achieved 3D object coordinates with photogrammetric bundle techniques is also presented.
Figure 1. Typical sequence (92 images) automatically oriented extracting the necessary image correspondences (18,500 3D points).
1. INTRODUCTION
The complexity and diversity of the image network geometry in
close-range applications, with wide baselines, convergent
images, illumination changes, moving objects, occlusions,
variations in resolution and overlap, makes the automated
identification of tie points more complex than in standard aerial
photogrammetry. Homologues image points are necessary for
structure and motion determination, as well as 3D modeling
purposes or Photosynth-like applications. In close-range
photogrammetry, commercial solutions for automated image
orientation and sparse 3D geometry reconstruction of
markerless sets of images are still pending. Some commercial
packages are available to automatically orient video sequences
(e.g. Boujou, 2D3 and MatchMover, RealViz), but they
generally work only with very short baselines and low
resolution images. Thus, there is a lack of commercial and
reliable software to automatically orient a set of unordered and
markerless images. In the literature there are some approaches
tailored to work in real-time and on large indoor or outdoor
environments. These methods, typically named SLAM, can also
use external information coming from exogenous sensors (e.g.
GPS or IMU) for better incremental motion estimation and
absolute geo-referencing (Davison et al., 2007; Agrawal and
Konolige, 2008; Pollefeys et al., 2008).
When no assumption or external information are employed, the
“Structure from Motion” (SfM) concept is the core method used
for the automated orientation of images and 3D sparse
reconstruction of scenes (Pollefeys and Van Gool, 2002). Nister
(2004) matches small subsets of images to one other and then
merge them for a complete 3D reconstruction in form of sparse
point clouds. Vergauwen and Van Gool (2006) developed a
SfM tool for Cultural Heritage applications (hosted now in a
web-based 3D reconstruction service). Recently the SfM
concept has made tremendous improvements, notwithstanding
the achievable 3D reconstructions are useful only for
visualization, object-based navigation, annotation transfer or
image browsing purposes. However, the automation of the
procedure has reached a significant maturity with the capability
to orient huge numbers of images. Two well-known packages
are Bundler (or its graphical version Photosynth) (Snavely et
al., 2008a) and Samantha (Farenzena et al., 2009). The former
is the implementation of the current state of the art for
sequential SfM applications, and it was also extended towards a
hierchical SfM approach based on a set of key-images (Snavely
et al., 2008b). The latter appears even faster because of the
introduction of a local bundle adjustment procedure.
In the photogrammetric community, some research solutions
capable of automatically orienting a set of markerless images
acquired with calibrated cameras were presented in Roncella et
al. (2005), Labe and Forstner (2006) and Remondino and Ressl
(2006). A rigorous bundle solution, coupled with the estimation
of the unknown parameters based on the Gauss-Markov model
of the Least Squares, provided an efficient, precise and reliable
solution in a functional and stochastic sense.
The paper presents a methodology for the automated
identification of image correspondences in a large variety of
image datasets (Figure 1). The proposed method, named with
