Vision-based uav guidance for autonomous landing with deep neural networks

Landing is one of the most critical phases of the flight during the operation of Unmanned Aerial Vehicles (UAVs). Even though flight control systems could perform automatic landing in nominal conditions, sensor failures in this phase might result in catastrophic crashes. In this study, we have trained an end-to-end Deep Learning (DL) model using the raw image inputs to estimate the relative heading angle of the aircraft with respect to the runway of the airport, when the connection with the corresponding sensor (beacon) is broken. To this end, we have used a closed-loop position trajectory following guidance and control system in order to train the network. The prediction performance of the network is shown for a number of unseen landing scenarios. Afterward, the estimated heading angle is fed into the guidance block in order to perform the autonomous landing using the angles predicted with the network. The main contribution of our work is to use an end-to-end architecture of the DL agent to estimate the observed state using raw image inputs, which increases the robustness with respect to sensor failures.