IEEE Robotics & Automation Magazine - December 2018 - 31

The correction is derived from an
analysis of the axial camera model for
underwater cameras, which is physical-
ly correct but computationally hard to
tackle. It can be shown how realistic
constraints on the distance of the cam-
era to the window can be exploited,
leading to an approach known as the
pinax model because it combines
aspects of a virtual pinhole model with
the projection function from the axial
camera model. The pinax model is not
only convenient because it allows in-air
calibration; it also outperforms standard
methods in terms of accuracy [12].

(a)

(b)

(c)

3-D Mapping and Object
Recognition
The data from the camera system with
its stereo setup are processed online to
generate dense 2.5-D point-clouds,
(d)
(e)
which are integrated into 3-D maps
(Figure 9). The well-known octree data
structure is used for this purpose. More Figure 9. When (a) the vehicle approaches (b) the mockup panel structure, (c) the
augmented reality marker can be used to aid the navigation to validate different
precisely, our implementation builds navigation methods. Among others, (d) a 3-D octomap is generated in real time, which is
upon the popular OctoMap library transmitted to the offshore control center. (e) Because the mockup panel structure is-as
[13], which is extended for differential in oil and gas operations-a priori known, perceived parts can be used to determine its
pose and project the known model in the scene to aid the execution of tasks. (Photos
update operations to support efficient, and screenshot courtesy of DexROV consortium.)
low-latency transmission of the 3-D
representation over the satellite link to
the onshore command center [14]. Furthermore, an efficient The mockup panel structure is equipped with augmented
strategy for underwater color updates is added.
reality markers (Figure 9), which can further aid the naviga-
Underwater images suffer from challenging light condi- tion and be used for validating the other navigation methods.
tions, especially wavelength-dependent attenuation and for- The lower-level image processing, i.e., rectification and
ward scattering and backscattering. When coloring the image enhancement, are important for the robust recogni-
octomap, a very simple but efficient strategy is used: because tion of the markers.
attenuation is wavelength and distance dependent, the
The 3-D octomap can be run with a 2-cm grid cell reso-
brightest measurement is used, which corresponds to the lution. This is well suited for autonomous operations like
closest and, hence, most accurate sample [14]. For substan- collision avoidance and path planning, and it is also suffi-
tial image enhancement-at much higher computational cient to give human operators in the onshore command
cost-a new variant of the dark channel prior for underwa- center an overview of the environment including unex-
ter vision is used [15]. Even though there are known adapta- pected obstacles and the terrain. To aid manipulation, we
tions of this method to underwater applications, further exploit the fact that, in many application cases, e.g., for oil
improvements are made by reformulating the problem. and gas operations, the structures to be dealt with are
Bright regions in the dark channel appear in outdoor images known. The detailed simulation framework employed for
in air on nonsky regions because of the backscattering, component and system validation [16] can be used in this
which is used in the original dark channel prior. In contrast context to provide a virtual visualization of the real under-
to other adaptations of the method to underwater applica- water operations (Figure 9), i.e., when the vision system of
tions, the estimation of the atmospheric light is adjusted, the ROV detects or tracks parts of the structure of interest,
which leads to clear improvements [15].
the structure's pose is transmitted to the onshore com-
The core navigation of the vehicle is based on the data of mand center and the known model can be projected into
the associated sensors connected to the vision compute bot- the scene. Furthermore, dedicated vision processes can
tle, i.e., the NavQuest DVL and the Xsens IMU. These data detect and localize crucial objects as a basis for (semi)
are fused in an extended Kalman filter and can be aided by autonomous manipulation. For example, the valves on the
the registration of the stereo scans and the tracking of objects mockup structure are detected and localized with an active
up to the level of full simultaneous localization and mapping. contours method having superellipse fitting. Their
december 2018

IEEE ROBOTICS & AUTOMATION MAGAZINE

IEEE Robotics & Automation Magazine - December 2018

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2018