Computer vision for improving the drone state estimate

Abstract

The drone industry is currently experiencing a fast-paced development which leads to the creation of multitude of various products. An emerging trend is the search of increased smartness and autonomy of the machines. A prerequisite for this quest of autonomy is however the need of having a robust and reliable state estimate over time. In this Master Thesis, we explore different possibilities of achieving this in real-time by using an on-board mounted camera in pair with other sensors for the Fleye, a drone developed by Aerobot. More specifically, we focus our attention on the Multi-State Constraints Kalman Filter for which we provide a detailed explanation and an implementation designed for the Fleye. The strength of this filter resides in its relatively good computational efficiency compared to its alternatives and in its ability to deal with some hardware uncertainties such as an approximative knowledge of the relative position of the different sensors. A method developed to generate synthetic data allowing to test the performance of visual-inertial odometry algorithms is presented in this work. Performance tests made on synthetic and experimental data show that the implemented filter is consistent but still requires further improvements in order to compete with current state-of-the-art solutions.