Towards Robust Visual-Controlled Flight of Single and Multiple UAVs in GPS-Denied Indoor Environments

Abstract

Having had its origins in the minds of science fiction authors, mobile robot hardware has become reality many years ago. However, most envisioned applications have yet remained fictional - a fact that is likely to be caused by the lack of sufficient perception systems. In particular, mobile robots need to be aware of their own location with respect to their environment at all times to act in a reasonable manner. Nevertheless, a promising application for mobile robots in the near future could be, e.g., search and rescue tasks on disaster sites. Here, small and agile flying robots are an ideal tool to effectively create an overview of the scene since they are largely unaffected by unstructured environments and blocked passageways.

In this respect, this thesis first explores the problem of ego-motion estimation for quadrotor Unmanned Aerial Vehicles (UAVs) based entirely on onboard sensing and processing hardware. To this end, cameras are an ideal choice as the major sensory modality. They are light, cheap, and provide a dense amount of information on the environment. While the literature provides camera-based algorithms to estimate and track the pose of UAVs over time, these solutions lack the robustness required for many real-world applications due to their inability to recover a loss of tracking fast. Therefore, in the first part of this thesis, a robust algorithm to estimate the velocity of a quadrotor UAV based on optical flow is presented. Additionally, the influence of the incorporated measurements from an Inertia Measurement Unit (IMU) on the precision of the velocity estimates is discussed and experimentally validated. Finally, we introduce a novel nonlinear observation scheme to recover the metric scale factor of the state estimate through fusion with acceleration measurements. This nonlinear model allows now to predict the convergence behavior of the presented filtering approach. All findings are experimentally evaluated, including the first presented human-controlled closed-loop flights based entirely on onboard velocity estimation.

In the second part of this thesis, we address the problem of collaborative multi robot operations based on onboard visual perception. For instances of a direct line-of-sight between the robots, we propose a distributed formation control based on ego-motion detection and visually detected bearing angles between the members of the formation. To overcome the limited field of view of real cameras, we add an artificial yaw-rotation to track robots that would be invisible to static cameras. Afterwards, without the need for direct visual detections, we present a novel contribution to the mutual localization problem. In particular, we demonstrate a precise global localization of a monocular camera with respect to a dense 3D map. To this end, we propose an iterative algorithm that aims to estimate the location of the camera for which the photometric error between a synthesized view of the dense map and the real camera image is minimal.