Towards High-Speed Localisation for Autonomous Drone Racing

Abstract

The ability to know the pose of a drone in a race track is a challenging task in Autonomous Drone Racing. However, to estimate the pose in real-time and at high-speed could be fundamental to lead an agile flight aiming to beat a human in a drone race. In this work, we present the architecture of a CNN to automatically estimates the drone’s pose relative to a gate in a race track. Due to the challenge in ADR, various proposals have been developed to address the problem of autonomous navigation, including those works where a global localisation approach has been used. Despite there are well-known solutions for global localisation such as visual odometry or visual SLAM, these methods may become expensive to be computed onboard. Motivated by the latter, we propose a CNN architecture based on the Posenet network, a work-oriented to perform camera relocalisation in real-time. Our contribution relies on the fact that we have modified and re-trained the Posenet network to adapt it to the context of relative localisation w.r.t. a gate in the track. The ultimate goal is to use our proposed localisation approach to tackle the autonomous navigation problem. We report a maximum speed of up to 100 fps in a low budget computer. Furthermore, seeking to test our approach in realistic scenarios, we have carried out experiments with small gates of 1 m of diameter under different light conditions.