A Temporal CNN-based Approach for Autonomous Drone Racing

Convolutional Neural Networks (CNN) and Deep Learning (DL) have become a popular tool to address all sorts of artificial intelligent challenges. The Autonomous Drone Racing is a challenge consisting of developing an autonomous drone capable of beating a human in a drone race, and DL is a tool that has been included in state of the art solutions to address this problem. Current works have proposed to use CNN and DL to detect the gates, whereas other works have proposed to use a CNN to obtain drone's control commands and a goal point, with all of these approaches using a single image as input. In this work we propose a CNN based on the well known pose-net network. Originally used for camera relocalisation, we propose to use pose-net to provide control commands to drive the drone towards and to cross the gate autonomously. In contrast to previous works, we also propose to use a temporal set of images as input for the network. In specific, we use 6 images captured every 166 milliseconds in one second to create a mosaic. The latter is used as input of the CNN to predict the control commands. We compare this proposed temporal approach against using a single image as input for the CNN. Our results, although in simulation, demonstrate that the using only our temporal approach is feasible, less noisy and more effective than the single image approach, enabling the drone to autonomously cross a set of gates placed randomly, and even under the scenario where the gate moves dynamically.