DEEP LEARNING OF VIRTUAL-BASED AERIAL IMAGES: INCREASING THE FIDELITY OF SERIOUS GAMES FOR LIVE TRAINING

摘要

The aim of rapidly reconstructing high-fidelity, Synthetic Natural Environments (SNEs) may benefit from a deep learning algorithm: this paper explores how deep learning on virtual, or synthetic, terrain assets of aerial imagery can support the process of quickly and effectively recreating lifelike SNEs for military training, including serious games. Namely, a deep learning algorithm was trained on small hills, or berms, from a SNE, derived from real-world geospatial data. In turn, the deep learning algorithm's level of classification was tested. Then, assets learned (i.e., classified) from the deep learning were transferred to a game engine for reconstruction. Ultimately, results suggest that deep learning will support automated population of high-fidelity SNEs. Additionally, we identify constraints and possible solutions when utilising the commercial game engine of Unity for dynamic terrain generation.