Computer Vision and Deep Learning in Autonomous Drones

Abstract

In this thesis we want to create a deep learning based object detection solution that is able to run locally on an autonomous drone. The goal of the drone is to herd a group of ground robots correctly within the International Aerial Robotics Competition (IARC). The main problem this thesis want to address is how to reliably detect these ground robots through cameras mounted on the drone. It all has to be performed in real time and in four different directions at the same time with limited computational power. The proposed solution is to first create a large dataset of detection examples using images recorded at the IARC 2016 competition and second to use the dataset to train the fastest and most accurate detection neural networks available like YOLO (You Only Look Once) and SSD (Single Shot Multibox Detection). The final result is a dataset of 6100 detection images with labels for each of the three different types of robot in the competition as well as several detection networks that are able to run in real-time onboard the drone with varying degree of accuracy.