Open Water Detection for Autonomous In-harbor Navigation Using a Classification Network

Autonomous navigation quay to quay is a goal for various surface vessel trades, from inland ferries to river transport and offshore services. Ability to navigate safely within a harbour or other confined waters is an essential step-stone towards this goal. This paper aims at creating a map of open water area that is available for safe navigation, given dynamic and static obstacles. Employing electro-optical sensors, the paper suggests open water detection using a classification convolutional neural network on context sensitive sub-partitioning of an image in a pyramid of smaller areas, combining the classifications in to a map of subareas containing open water. A salient feature of this approach is the ease of annotation and ease of creating a large amount of annotated images that is needed for machine learning. Following classification of sub-areas, camera images are transformed to bird’s view by projective geometry methods to enable planning of feasible paths for navigation. This new approach is validated on data from sea trials in Danish waters.