Autonomous robotic exploration of coral reefs using a visual attention-driven strategy for detecting and tracking regions of interest

Visual-based autonomous robotic exploration of unstructured and highly dynamic environments is a complex task. We present an approach to carry out an attention-driven exploration of underwater environments. This work is aimed to grant autonomy to an exploring agent in terms of deciding where to move in function of relevant visual information. This way we could obtain close video-observations of regions of interest in coral reefs in order to diagnose disease or physical damage. Our approach first detects relevant points from the images the vehicle is capturing, this is achieved by utilizing a visual attention model adapted to work on underwater scenes. Then, a particular region of interest can be quickly and robustly tracked by using superpixel descriptors, which are associated to each of the relevant points. The tracking continues as long as it results interesting for the visual attention algorithm. The field experimental results show the effectiveness of the proposed approach.