Low Altitude Aerial Scene Synthesis Using Generative Adversarial Networks for Autonomous Natural Resource Management

摘要

Deep neural networks are currently the best solution for aerial scene interpretation for Unmanned Aerial Vehicle (UAV) based remote sensing. A problem faced by the deep neural networks is that the deep models require significantly large training datasets which should cover almost all of the scenarios. Gathering these datasets is usually very time consuming and expensive. In this paper, data augmentation and generative adversarial network are used for autonomous synthesis of low altitude aerial scenes for creating a training dataset for deep low altitude aerial video interpretation. The proposed system is evaluated using a real world scenario of road following under foliage in a jungle and the experimental results show that the proposed framework is capable of producing high accuracy training datasets for UAV vision system in natural resource management scenarios.