DORA@CRAS

In the context of the MSc thesis entitled "A Machine Learning Approach for Predicting Docking-based Structures", the dataset denominated DOcking-based stRuctures dAtaset (DORA) was created to provide information for the training of a Deep Learning Network which aimed to detect any docking platforms present in unstructured maritime environments. The dataset was acquired in 3D simulator Gazebo. Different types of docking-based structures were placed in a simulated maritime environment along with other types of objects such as floating buoys and boats.

A model of an Autonomous Surface Vehicle (ASV), with a set of incorporated sensors (LiDAR, left and right cameras, GPS, IMU) was sequentially positioned in full circles of increasing distance around the dock. To further increase variety in the dataset, the orientation of the vehicle also varied, by modifying the Euler angles roll, pitch and yaw. This collection process resulted in a dataset of circa 34000 instances, 19000 of which contained a docking plaform and 15000 did not.