Recently, as UAVs (unmanned aerial vehicles) have become smaller and higher-performance, they play a very important role in the Internet of Things (IoT). Especially, UAVs are currently used not only in military fields but also in various private sectors such as IT, agriculture, logistics, construction, etc. The range is further expected to increase. Drone-related techniques need to evolve along with this change. In particular, there is a need for the development of an autonomous system in which a drone can determine and accomplish its mission even in the absence of remote control from a GCS (Ground Control Station). Responding to such requirements, there have been various studies and algorithms developed for autonomous flight systems. Especially, many ML-based (Machine-Learning-based) methods have been proposed for autonomous path finding. Unlike other studies, the proposed mechanism could enable autonomous drone path finding over a large target area without size limitations, one of the challenges of ML-based autonomous flight or driving in the real world. Specifically, we devised Multi-Layer HVIN (Hierarchical VIN) methods that increase the area applicable to autonomous flight by overlaying multiple layers. To further improve this, we developed Fisheye HVIN, which applied an adaptive map compression ratio according to the drone's location. We also built an autonomous flight training and verification platform. Through the proposed simulation platform, it is possible to train ML-based path planning algorithms in a realistic environment that takes into account the physical characteristics of UAV movements.
Keywords: Fisheye; IoTs; UAVs; VIN; autonomous flight; machine-learning.