Ship Hull Cleaning Robots

Abstract

Fouling (algae, slime, and barnacles) on the hull of large cargo vessels is undesirable because it increases their frictional drag, resulting in an increased fuel consumption. As a solution, Fleet Cleaner introduced a ship hull cleaning robot that maneuvers on the hull, using powered wheels and magnets. The robot is controlled by a human operator whose main responsibility is to ensure safe navigation over the hull, by avoiding sharply curved surfaces and irregularities on the hull. The operators are subjected to high levels of mental fatigue because constant vigilance is required. Fleet Cleaner aims to alleviate the operators from demanding missions, by shifting towards a semi-autonomous mode of operation. This thesis contributes to this transition, by developing an obstacle detection framework that is tailored to the primary imaging device of the robot: a forward looking sonar (FLS). This device provides an acoustic image that is characterized by a low signal to noise ratio, a nonuniform illumination, and its sensitivity to viewpoint changes. To the best of the author’s knowledge, no research has been reported about FLS-based obstacle detection for the working environment of ship hull cleaning robots. Therefore, this thesis contributes to both the Fleet Cleaner robot and the research field of acoustic imaging. The thesis scope is delimited by defining two objectives.
Firstly, an algorithm is proposed that detects the line parameters of the bilge, i.e., the rounded transition between the side and the bottom of a ship. Secondly, an algorithm is proposed that detects the boundaries of all obstacles that feature one distinct attribute: a grating pattern.

The design process is guided by four working steps: task clarification, conceptual design, implementation, and a proof of concept (PoC). The task clarification phase defines metrics that measure the detection accuracy of both algorithms, with respect to hand-drawn ground truths. The conceptual design phase establishes candidate solutions for both objectives. For bilge detection, two edge detectors and three line detectors are selected. For grating detection, a supervised approach is preferred, resulting in six texture features, one segmentation algorithm, and two classifiers. The last two working steps employ the conceptual solutions on recorded FLS images, such that the detection accuracy can be computed. The detection accuracy is used to tune the parameters of all algorithms in the implementation phase and to
select the final solution in the PoC.

The PoC of the bilge detector demonstrates that a sufficiently high accuracy was attained by the Radon transformation (line detector), in combination with a first-order steerable filter (edge detector). The PoC of the grating detector shows that highest accuracy was obtained for a fused set of texture features (local binary patterns and Haralick’s features), in combination with the support vector machine classifier. However, the achieved accuracy was not sufficiently high. It is believed that the lacking accuracy is due to the limited amount of training data that was available and the deficiency of the employed segmentation algorithm. By resolving these issues, and by repeating the four working steps for other obstacles on the hull, it is believed that a complete obstacle detection framework can be acquired in the future.