Vision based navigation for micro helicopters

In this dissertation, we study the issues of vehicle state estimation and sensor self-calibration which arise while navigating a micro helicopter in large and initially unknown environments. While proper vehicle state estimation allows long-term navigation, sensor self-calibration renders the vehicle a power-onand-go system without the need of prior and offline calibration steps. Due to the inherent payload limitation of airborne vehicles like, and especially, a micro helicopter and the restricted availability of global localization information in indoor or urban environments, we focus on vision based methods in order to achieve our goals. Despite this focus on processing visual cues, we analyze a general and modular approach which allows the use of a variety of different sensor types. In vision based methods, the camera needs special attention. In contrast to other sensors, vision sensors typically yield vast information that needs complex strategies to permit use in real-time and on computationally constrained platforms. We show that our map-based visual odometry strategy derived from a state-of-the-art structure-from-motion framework is particularly suitable for locally stable, pose controlled flight. Issues concerning drifts and robustness are analyzed and discussed with respect to the original framework. A map-less strategy based on optical flow and inertial constraints is presented in order to mitigate other map-related issues such as map losses and initialization procedures. While in the map-based approach the camera acts as a real-time pose sensor, in the map-less approach it acts as a realtime body-velocity sensor capable of velocity-control the airborne vehicle. We show that both approaches can be combined in a unifying and complementary framework to mitigate each others weaknesses. In all cases, the algorithms are capable of running in real-time on a computationally constrained platform. For aerial vehicle navigation we study a statistical and modular sensorfusion strategy. We focus on the optimal state estimation of the vehicle to avoid additional, computationally heavy control approaches. In particular, we discuss the metric recovery of an arbitrarily scaled pose or body-velocity