Object localization and identification for autonomous operation of surface marine vehicles

Abstract

A method for autonomous navigation of surface marine vehicles is developed A camera video stream is utilized as input to achieve object localization and identification by application of state-of-the-art Machine Learning algorithms. In particular, deep Convolutional Neural Networks are first trained offline using a collection of images of possible objects to be encountered (navy ships, sail boats, power boats, buoys, bridges, etc.). The trained network applied to new images returns real-time classification predictions with more than 93% accuracy. This information, along with distance and heading relative to the objects taken from the calibrated camera, allows for the precise determination of vehicle position with respect to its surrounding environment and is used to compute safe maneuvering and path planning strategy that conforms to the established marine navigation rules. These algorithms can be used in association with existing tools, such as LiDAR and GPS, to enable a completely autonomous marine vehicle.