Energy-Aware Design of Vision-Based Autonomous Tracking and Landing of a UAV

In this paper, we present the design and evaluation of a vision-based algorithm for autonomous tracking and landing on a moving platform in varying environmental conditions. We use an energy-aware approach, where the design of the algorithm is based on an evaluation of the energy consumption and Quality of Service (QoS) of each computational component. We evaluate our approach with an agricultural use case where a moving platform is tracked using a landing marker and the YOLOv3-tiny CNN is used to detect ground-based hazards. We perform all computations onboard using an NVIDIA Jetson Nano and analyse the impact on the flight time by profiling the energy consumption of the marker detection and the CNN. Experiments are conducted in Gazebo simulation using an energy modeling tool to measure the computational energy cost as a function of QoS. We test the energy efficiency and robustness of our system in various dynamic wind disturbances. We show that the marker detection algorithm can be run at the highest QoS with only a marginal energy overhead whereas adapting the QoS level of CNN results in a considerable power saving. The power saving is significant for a system executing on a fixed-wing UAV.