621
MULTI-IMAGE MATCHING: AN “OLD AND NEW” PHOTOGRAMMETRIC ANSWER
TO LIDAR TECHNIQUES
F. Nex a *, F. Rinaudo a
a Dept, of Land, Environment and Geo-engineering (DITAG), Politecnico di Torino, Corso Duca degli Abruzzi 24,
Torino, Italy, (francesco.nex, fulvio.rinaudo)@polito.it
Commission V, Working Group V/3
KEY WORDS: Close Range Photogrammetry, Digital, Matching, Segmentation, Software, Accuracy, Cultural Heritage
ABSTRACT:
Over the last decade, LIDAR techniques have replaced traditional photogrammetric techniques in many applications because of
their speed in point cloud generation. However, these laser scanning techniques have non-negligible limits and, for this reason,
many researchers have decided to focus on improving the performances of matching technique in order to generate dense point
clouds from images.The first tests carried out at the Politecnico di Torino on the first fully-automated multi-image matching
commercial software, the ZScan Menci Software, are described in this paper. This instrument was first devised to allow
inexperienced users to generate very dense point clouds from image triplets; a customized calibrated bar (0,90 m length) is used for
image acquisition. Recently a new version of ZScan has been created in order to elaborate triplets of oriented aerial images and
generate DSM: the first results obtained in this way are presented in this paper. Several tests have been performed on the ZScan
performances analysing different geometrical configurations (base-to-height ratio) and textures. The evaluation of the geometric
precision obtained by this software in point cloud generation may help to understand which performances can be achieved with a
fully automated multi-image matching. The evaluation concerns what the most critical aspects of these techniques are and what
improvements will be possible in the future. Furthermore a possible new research project is described which has the aim of
transferring useful information about breakline location from images to point clouds in order to derive automatically the
segmentation algorithms.
1. INTRODUCTION
Over the last decade, LIDAR techniques have replaced
traditional photogrammetric techniques in many applications,
because of their speed in point cloud generation and the
percentage of acceptable points; traditional image matching,
based on the use of stereopairs, can acquire no more than 80%
of the possible points (the holes have to be manually
completed).
However, these laser scanning techniques have non-negligible
limits due to the impossibility of directly obtaining radiometric
information and the exact position of object breaklines. For this
reason, most LIDAR surveys are integrated by digital image
acquisition. Digital photogrammetry directly associates a
radiometric information to the acquired points and the use of a
stereopair allows a manual survey of the breaklines when
automatic algorithms fail. Then, other limits for LIDAR
terrestrial applications are the weight and size of the
instruments, the limited range of applications (especially for
environmental applications) and last but not least the cost.
Finally, when point clouds are produced, segmentation and
classification procedures have to be applied in order to correctly
interpret and model the surveyed object. The acquired
experience shows that automatic algorithms are not ready to
offer correct solutions without direct human help or without an
image interpretation.
For these reasons, many researchers have decided to focus on
improving the performances of matching techniques in order to
generate dense point clouds from only images.
2. MATCHING ALGORITHMS
As is well known, these techniques were devised more than
twenty years ago and are nowadays subdivided into Area Based
Matching (ABM) and Feature Based Matching (FBM).
Since then, the improvement of these techniques has been
ongoing with research into new algorithms that are capable of
improving their efficiency and precision. This research has led
to matching algorithms which have become more complete
(Gruen, 1985) and improved their effectiveness, imposing new
constraints, such as collinearity conditions (Gruen, Baltsavias,
1988), and integrating the surface reconstruction into the
process. Furthermore, these techniques now allow more than
two images to be handled and simultaneous determination of all
matches of a point: these features help overcome the problems
of mismatches thus increasing the precision achieved (Forstner,
1998).
These techniques have recently been improved by other authors
who have adapted the algorithms to aerial (Gruen, Zhang, 2004)
and terrestrial photogrammetry (El-Hakim, et al.2007;
Remondino, 2007), partly overcoming some of the matching
problems (wide baselines between images, illumination changes,
repeated pattern and occlusions). An important aspect of these
Corresponding author.