Dynamic Positioning of an Underwater Vehicle using Monocular Vision-Based Object Detection with Machine Learning

Sonar and camera are two widely used sensors in the underwater segment. Moreover, optical based systems provide higher spatial and temporal resolution than their acoustic counterpart. In this paper, a dynamic positioning system for a small-class ROV relative to an object detected by a monocular camera will be presented. The object detection will be performed by the state-of-the-art object detector YOLOv3 trained on a dataset representing the relevant known object. In order to perform dynamic positioning based on 2D images, a scaling will be used to extract the spatial features from the images. The entire system is able to perform in real-time, which is essential for a dynamic positioning procedure.