Convolutional Neural Networks for Geo-Localisation with a Single Aerial Image

Nowadays, Unmanned Aerial Vehicles (UAVs) navigating outdoors rely heavily on GPS for localisation and autonomous flight or applications for aerial photography re-cording with a GPS coordinate. However, GPS may fail or become unreliable, thus compromising the flight mission. Motivated by this scenario, in this work, we present a study on the use of popular Convolutional Neural Networks (CNN) to address the problem of geo-localisation from a single aerial image. We compare CNN-based architectures from the state-of-the-art, and introduce a compact architecture to speed up the inference process without affecting the estimation error. For our experiments, aerial images were recorded with a monocular camera onboard a UAV, flying outdoors with a height between 20 to 25 metres. On average, our compact network achieves a minimum estimation error of 2.8 metres and a maximum of 6.1 metres, which is comparable to the performance of other networks in the state-of-the-art. However, our network achieves on average an operation frequency of 103 fps versus 69 fps achieved by the fastest network in the comparison analysis. These results are promising since such speed would enable fast geo-localisation with cameras capturing images at those frame rates, which are useful for obtaining neater images than with conventional cameras working at 30 fps.