Underwater photogrammetry
26
Hydrographische Nachrichten
Fig. 3: Original image (left) and enhanced Image using LAB algorithm (right)
sated by image distortion parameters (Menna et
al. 2016; Nocerino et al. 2016). Otherwise, the ray
path could be modelled explicitly by applying ray
tracing approaches. This however, would reguire
a specific bundle adjustment solution, eliminating
the option of using standard structure-from-mo-
tion (SfM) software, as it is commercially available
to users from administration and industry.
Images were acguired at a frame rate of 20 Hz.
However, in order to reduce computational and
memory effort, images were analysed at 2 Hz. At
approximately 0.5 m/s lateral movement speed,
an acguisition distance of 1 m and a ground sam
pling distance (GSD) of 1 mm has been achieved,
this leads to an average overlap of 87 % in hori
zontal and 79 % in vertical direction. This is con
sidered enough overlap to be able to robustly
identify identical features over several images in
a seguence. The total survey time, including dive
time and time to locate the wreck, was about
15 min. of which 7.5 min. consisted in the imagery
acguisition time.
The Baltic Sea has a high turbidity and therefore
does not provide very good visibility conditions.
In order to improve matching results and colour
correctness, several image enhancement meth
ods, as proposed in Mangeruga et al. (2018) were
compared. By far the best results were achieved,
using the LAB enhancement algorithm, proposed
by Bianco et al. (2015). Fig._3 displays a wreck fea
ture viewed in the original image and the same
feature viewed in an image enhanced by the LAB
algorithm. The original image is obviously biased
towards green, which distorts the wreck feature
and reduces contrast. The enhanced image on
the other hand still has a green background but
the wreck feature is more distinguishable from the
background. This leads to improved matching re
sults.
Using the enhanced imagery, photogrammetric
analyses were performed using structure-from-
motion (SfM) processing methods. SfM technigues
(e.g.Snavelyetal.2006; Furukawa and Ponce 2010)
generate 3D representations from 2D image se-
guences without initial information. Feature points
are extracted from the images and matched,
employing robust estimation technigues such
as RANSAC (random sample consensus; Fischler
and Bolles 1981). Using these corresponding im
age points in multiple images, bundle adjustment
was performed using a self-calibration approach.
From this method, the interior orientation was cal
culated using distortion parameters according to
Brown (1971), i.e. principal distance, principal point,
radial-symmetric and decentering distortion, and
affinity and shear. Simultaneously, values for exte
rior orientation (6DOF position of camera in object
space) and 3D coordinates of the object points
were estimated. The aforementioned steps were
performed, using Agisoft Metashape, a widely
used SfM software that has been proven to be ro
bust in underwater photogrammetry (Mangeruga
et al. 2018). Unfortunately, due to the commercial
aspect of the software, no detailed insights are
provided about the algorithms used for orienta
tion and subseguent dense image matching (Re-
mondino et al. 2013). For this image bundle, 606
images from the starboard side were aligned with
Fig. 4: RGB-coloured sparse point cloud and camera trajectory (red)
