Fleet Cleaner B.V. is a Dutch company which deploys a wheeled hull-fixed Remotely Operated Vehicle (ROV) to clean biofouling off the Vertical Sides (VSs) and the Flat Bottom (FB) of the hull of seagoing cargo vessels, while berthed at the harbour. To ensure maximal cleaning coverage, accurate localization to keep track of the robot trajectory on the ship hull is essential, but also challenging. Contrary to the VS case, drift-free localization from motion sensors is not applicable on the FB. Instead, the current system relies heavily on operator input, who in turn relies on features (such as the quay wall, a line feature) appearing on imagery from a Forward Looking Sonar (FLS) to determine the robot pose. Human inattentiveness caused by the high workload of cleaning operations makes this method error-prone and sensitive to drift. In the future, Fleet Cleaner aims to develop an autonomous robot to relieve the operators from the task of localization and improve cleaning performance. As no previous work was performed on this matter, this thesis is a contribution to this vision, by developing a sonar-based framework enabling drift-free localization of a ship hull cleaning robot on FBs. Before the conceptual phase, two objectives are formulated to achieve the thesis goal. The first is to provide a sonar-based algorithm measuring the line parameters of the quay wall, as lateral and heading reference. The second objective seeks to fuse the FLS-based measurements with motion sensors to achieve drift-free localization. From there, the operating conditions relevant to the systems are defined, and both functional and performance requirements are formulated, where the latter define metrics that measure the accuracy of the systems w.r.t. manually defined ground truths.
For the first objective, three candidate solutions for tracking the quay wall line parameters are selected. These are first implemented with recorded data in an iterative tuning procedure, then validated where the Total Least-Squares (TLS)-based line tracker with Gaussian filtering arises as a proof of concept. Although the algorithm is fast enough to function for each incoming sonar image, it does not fulfill the requirements for accuracy and robustness, a problem solvable by tuning with better data and the use of filters. For the second objective, also three sensor fusion algorithms are selected. Analysis of their properties with regards to performance revealed the Extended Kalman Filter (EKF) as the most suitable solution, where tuning was performed heuristically. Validation tests again showed partial success, with fast computation times and greatly improved but still insufficient accuracy of the FLS measurements, although localization drift was reduced within the required limits. Mainly, the poor predictions from auxiliary sensors degrade performance, but also the imperfect nature of manually defined ground truths exaggerate the errors. To improve and complete the localization system, it is recommended to implement trajectory optimization for an improved fit of estimated poses with sensory data, and Simultaneous Localization And Mapping (SLAM) based on weld line detection to reduce longitudinal drift.