Vision-based UAV Safe Landing exploiting Lightweight Deep Neural Networks
Charalampos Symeonidis, Efstratios Kakaletsis, Ioannis Mademlis, Nikos Nikolaidis, Anastasios Tefas, Ioannis Pitas
Abstract
Recent advances in artificial intelligence, control and sensing technologies have facilitated the development of autonomous Unmanned Aerial Vehicles (UAVs, or drones) able to self-navigate in various settings. Although these technologies have already entered a mature stage, ensuring flight safety in crowded areas or performing an emergency landing in case of malfunctions, while adhering to relevant legislation, is generally treated as an afterthought when designing autonomous UAV platforms for unstructured environments. This paper proposes a UAV safe landing navigation pipeline that relies on lightweight computer vision modules, able to be executed on the limited computational resources on-board a typical UAV. Pre-trained Deep Neural Networks (DNNs) are mainly employed as the underlying building blocks, since deep learning has made a major impact on robotic perception by drastically improving the performance of relevant tasks, such as object detection or tracking, semantic image segmentation, etc. Evaluation of the proposed pipeline on a simulated environment indicates highly favorable results.