Autonomous Visual-Inertial Navigation for Quadrotor MAVs

Abstract

The aim of this thesis is to develop a system which would enable a quadrotor MAV (Micro Aerial Vehicle) to estimate its position and orientation and to autonomously navigate in unknown environments using vision as the primary source of information. To navigate in three-dimensional space, an autonomous MAV should not only possess knowledge of its current position and orientation (pose for short), but also of the world around it. While the former can be obtained using a GPS for large outdoor environments and the latter can be provided as a map, a truly autonomous navigation system should enable an MAV to infer its pose in indoor, GPS-denied environments using only the on-board sensors. While images from a camera are rich in data, they are devoid of any depth information. Extracting depth information from a single camera therefore requires the presence of reference objects with known geometry such as artificial fiducial markers in the field of view, or state-of-the-art monocular structure from motion techniques. In this thesis, we will study and develop solutions for environments which the MAV possesses no a priori information. We will present our modular approach to the problem of autonomous navigation which comprises of three separate blocks. The first block uses a state-of-the-art monocular simultaneous localization and mapping algorithm to transform the camera into a real time pose sensor. The second block uses an extended Kalman filter to refine the pose information from the camera by fusing it with data from the MAV’s onboard sensors. The third block uses a proportionalintegral- derivative controller to generate control commands for the MAV. We implemented our system on a commercially available quadrotor MAV and tested it in real world scenarios. The accuracy of our system will be compared against a highly accurate motion capture system and the findings will be presented/analyzed.