Two-step gradient-based reinforcement learning for underwater robotics behavior learning
Robotics and Autonomous Systems
| Hi-index | 0.00 |
The surveillance and inspection of underwater installations such as telecommunication cables are currently carried out by trained operators who, from the surface, guide a remotely operated vehicle (ROV) with cameras mounted over it. This manual visual control is, however, a very tedious job that tends to fail if the operator loses concentration. This paper describes a tracking system for underwater narrow telecommunication cables, the main objective of which is to allow an autonomous underwater vehicle to video-document the whole length of a cable. The approach is based on particle filters (PFs) because of their natural ability to model multi-dimensional multi-modal PDFs, which allows handling in a more appropriate way the ambiguities that naturally arise from undersea environments. In effect, despite the special visual features that artificial objects present, which allow distinguishing them in natural scenarios such as the seabed, distracting background such as rocks or algae growing on top and nearby cables, complicate the detection and tracking and give rise to ambiguities when rocks or marine growth form shapes and textures that resemble the cable. Apart from the different models that a PF requires, the paper also describes a set of added features, which successfully compensate some large errors in the cable pose estimation when the non-enhanced tracker is applied. Extensive experimental results over a test set of more than 10,000 frames, for which a ground truth has been manually generated, have shown the usefulness of the solution proposed. Besides, results for a set of six video sequences accounting for almost 150,000 frames and around one hour and a half of successful continuous video tracking are also discussed. All those images come from inspection runs captured by ROVs navigating over real telecommunication undersea cables.