Drone safe to launch system using machine learning

Abstract

Svenska Sjöräddningssällskapet has partnered with Infotiv to develop a prototype drone and launcher system for search-and-rescue missions at sea. In this thesis project, we investigate the possibility of automating parts of the launch sequence of a seaborne surveillance drone. The main goal is to train a neural network model to use a camera feed to determine if it is safe or not to launch the drone in a given direction, and then integrate this solution with the graphic user interface through an application programming interface. Monocular depth estimation using transfer learning and the KITTI data set is evaluated. The KITTI data set does not contain maritime scenery, leading to an unsatisfactory monocular depth estimation model. U-net and convolutional neural network models are trained on the MaSTr1325 data set, which contains semantically segmented maritime imagery. We collect additional data for the semantic segmentation models and create a post-processing step that evaluates if it is safe to launch or not. These models yielded satisfactory results, and the convolutional neural network will be used by the drone operator as an extra safety measure during launch.