Aerial Geo-Localisation for MAVs using PoseNet

The Global Position System (GPS) has become an essential sensor for drones. Autonomous flight in outdoor areas is possible thanks to the use of GPS that enables the drone to obtain its position in latitude and longitude coordinates. However, GPS may become unreliable when the drone flies in environments where the signal may get occluded. Malicious attacks may also compromise the GPS signal, aiming at blocking the signal or replacing it with spurious data. Motivated by these scenarios, we present preliminary results of a methodology aimed at estimating the GPS position of a drone using Convolutional Neural Networks (CNN) and a learning-based strategy. For the latter, we have adopted the PoseNet CNN architecture, originally proposed to address the relocalisation or kidnapping camera problem for facing forward cameras. First we trained PoseNet with a set of aerial images captured with an on-board camera, providing X, Y and Z coordinates as labels, which are obtained from converting GPS coordinates into metres for X and Y, and using the altimeter for Z. Then we perform validation flights where the vehicle follows a different trajectory to that used for collecting the training datasets. Even when the terrain includes bushes and repetitive texture, the CNN returns predictions with an error around the 2.5 metres and a processing speed of 15 milliseconds on average. We argue that a system such as this could be used as an emergency option to return the drone to home in the event of GPS failure. To our knowledge, this is the first time PoseNet is tested to address the problem of geo-localisation of aerial images.