Bioinspired vision-only UAV attitude rate estimation using machine learning

This paper presents a bioinspired system for attitude rate estimation using visual sensors for aerial vehicles. The sensorial system consists of three small low-resolution cameras (10x8 pixels), and is based on insect ocelli, a set of three simple eyes related to flight stabilization. Most previous approaches inspired by the ocellar system use model-based techniques and consider different assumptions, like known light source direction. Here, a learning approach is employed, using Artificial Neural Networks, in which the system is trained to recover the angular rates in different illumination scenarios with unknown light source direction. We present a study using real data in an indoor setting, in which we evaluate different network architectures and inputs.