Generating Reality-Analogous Datasets for Autonomous UAV Navigation using Digital Twin Areas

In order for autonomously navigating Unmanned Air Vehicles(UAVs) to be implemented in day-to-day life, proof of safe operation will be necessary for all realistic navigation scenarios. For Deep Learning powered navigation protocols, this requirement is challenging to fulfil as the performance of a network is impacted by how much the test case deviates from data that the network was trained on. Though networks can generalise to manage multiple scenarios in the same task, they require additional data representing those cases which can be costly to gather. In this work, a solution to this data acquisition problem is suggested by way of the implementation of a visually realistic, yet artificial, simulated dataset. A method is presented for the creation of a 'Digital Twin Area' inside of a modern high fidelity game engine using 3D scanned models of physical locations, and a realistic dataset of each area is created to showcase this concept.