Evaluation of localization performance of different algorithms and sensors in the context of SLAM for indoor navigation of unmanned aerial systems

Abstract

Simultaneous Localizing and Mapping (SLAM) performing systems provides mobile robots with pose estimations and are indispensable when striving towards true autonomy. An aerial mobile robot, known as an Unmanned Aerial Vehicle (UAV), sets particularly high requirements to a possible SLAM system's accuracy, since collisions are intolerable. In many cases, increased accuracy comes at the price of increased computational power and higher weight of the environment perceiving sensors. Both are limited in UAVs because of the limitations in onboard energy storage and payload weight. If the UAV operates in indoor environments, the SLAM system has to function without the use of any satellite navigation. This thesis provides an overview of publicly available state of the art SLAM systems and an evaluation of their performances on the indoor sequences of the Viral dataset from the NTU. This dataset was recorded with two 3D-LiDARS and a stereo-camera and provides a ground truth, recorded with a Motion Capture System (MOCAP). On this dataset, it was possible to show that the LiDAR based SLAM systems, considered in this thesis, performed the position estimation with higher accuracy then the systems only using visual information. Especially, FAST-LIO was capable of very accurate positional estimation, while VINS-Fusion was the most accurate visual based system.