Underwater photogrammetry
24
Hydrographische Nachrichten
ronments. Moreover, sounding measurement un
certainties not exceeding the IHO specification for
Special Order Surveys (IHO 2008) provide a geo
metrically realistic representation of these objects.
Unfortunately, the range and depression angle
between the MBES and the object Inevitably af
fect the sounding measurement accuracy and,
more importantly In this use case, the ability to
ensonify all parts of the object, especially on com
plex objects such as wrecks. A reliable assessment
of the state of wrecks solely based on sounding
measurements is therefore guestionable. Profes
sional divers and camera-eguipped ROV provide
the necessary close range in-situ inspection. How
ever, due to poor visibility conditions, this inspec
tion Is much localised. A thorough assessment of
the state of large objects (e.g. wrecks) is necessarily
the synthesis of many such localised inspections,
which Is Inherently subjective and error-prone.
Series of close-range underwater images col
lected by a camera-eguipped ROV over wrecks,
whilst still localised, are much easier to amalgam
ate to portray the whole structure of a wreck.
Cameras are portable and small, thus being eas
ily mountable on ROVs or attached to divers. Data
in occluded areas is collected by simply circling
the object. Photogrammetric analysis carried out
on these images results in geometrically-correct
three-dimensional point clouds characterised by
high spatial and temporal resolution as well as col
our information (Luhmann et al. 2020). When the
imagery-based 3D point cloud is co-registered to
the MBES soundings, the result is a fused data set
characterised by higher data density and addition
al attribution (i.e. colour information). Moreover,
targeting the ROV-lmages to the areas with low
sounding densities (I.e. occluded areas) allows for
an improved assessment of the state of a wreck.
Finally, being fixed to a common terrestrial refer
ence system, the fused data set is easily transfer-
rable to the spatial data infrastructure of maritime
administrations.
Fusing MBES data with information generated
from cameras Is thus highly desirable. However,
underwater imagery suffers from many degrading
and altering effects. This Includes multimedia ef
fects, as light travels through air, glass and water
and thus, according to Snell's law, the ray is refract
ed twice at the interfaces. This, by definition ren
ders the pinhole model invalid if no constructional
corrections are employed. Strict modelling of the
ray path has been developed, e.g. by Kotowski
(1988), Maas (1995) and Jordt-Sedlazeck and Koch
(2012). Several authors on the other hand found
that when the camera is positioned close to a flat
glass interface and oriented perpendicularly to It,
refraction effects can be compensated by stand
ard lens correction functions, as in Brown (1971),
and strict modelling is only decisive In applications
where highest accuracy is demanded (Kotowski
1988; Przybllla et al. 1990; Shortls 2015; Kahmen
et al. 2019). Furthermore, the entrance pupil of a
camera lens can be adjusted with the centre of a
hemispherical dome port. This accounts for Image
degradation, and possible residual errors are com
pensated by standard lens correction functions
(Menna et al. 2016). Furthermore, optical degrada
tion from wavelength dependent light absorption,
chromatic aberration or dispersion reduces image
guality. This results in Images with low contrast,
colourcast, blur and haze (Wang et al. 2019). To ac
count for these, several image enhancement and
restoration algorithms have been developed over
the years. These take the actual Image formation
model into account (e.g. Akkaynak and Trelbltz
2019) or employ appropriate Image processing
tools, such as histogram stretching, white bal
ance shift or gamma stretch to increase contrast,
decrease colour cast, etc. In Blanco et al. (2015),
the FAB method Is Introduced. Here, using a grey-
world assumption, the chromatic component of
the FAB colour space Is shifted towards the white
point and the luminance component Is enhanced
by histogram stretching and cut off. Thus, the
method belongs to the latter kind of algorithms.
Mangeruga et al. (2018) compared five state-of-
the-art Image enhancement algorithms for under
water photogrammetry and provided a metric for
benchmarking these. It was concluded that for 3D
reconstruction purposes, images enhanced with
the FAB algorithm or the original Images perform
best on their data sets.
As photogrammetry cannot provide absolute
positioning, further sensor data has to be com
bined with the Imagery, providing a georeferenced
position and point cloud. Furthermore, Imagery
can be used for online algorithms, solving posi
tioning and mapping of the environment in real
time (simultaneous localisation and mapping,
SEAM). These algorithms suffer from drift, as they
can only take a certain amount of data points and
positions into account, in order to not overflow
the memory and reduce computational complex
ity. Hence, the system drifts with time and distance
travelled, thus requiring additional information for
applications demanding high accuracy (Durrant-
Whyte and Bailey 2006). Originating from the ro
botics community, SEAM is a broad research field
and several methods have been proposed In re
cent years, using various kinds of sensors. State-of-
the-art algorithms are proposed by Mur-Artal and
Tardos (2017) or Engel et al. (2015), creating sparse
or seml-dense point clouds respectively. These
algorithms are capable, provided pre-callbrated
cameras, of computing point clouds and provide
localisation within the point clouds In real-time,
depending on the image resolution. Furthermore,
they automatically identify revisited areas and
compute so-called loop-closures, i.e. creating con
sistency between non-seguentlal parts of the data