Abstract
Pose estimation is of paramount importance for flight control as well as localization and navigation of Unmanned Aerial Vehicles (UAVs) to enable autonomous operations. The wide availability of consumer UAVs and its high autonomous capabilities are a testament to the maturity of this field. However, most platforms are translationdominant systems with rotations comprising a small part of the overall motion. As such, vision-based algorithms generally treat rotations as a subproblem of the overall motion scheme. Free rotors are a class of nature-inspired UAVs where their design and flight dynamics are inspired by the falling samara seed. With a constantly rotating body frame, free rotors introduce some unique challenges for visual perception where large rotations and small translations are tightly coupled in an image frame. This dissertation addresses the research problem of vision-based pose estimation on a free rotor. In total four technical contributions are made. Firstly, an algorithm that capitalizes on the translation invariant nature of the frequency domain for estimation of rotation. This method decouples the rotation and translation estimation into a two-step problem. By selecting keyframes in an image stream, the algorithm also outputs a stable translation-dominant image stream suitable for operator view. Secondly, the problem of image eccentricity is studied to boost the stable image stream frame rate. Since free rotors rotate about its center of mass for flight. Small changes in mass distribution are sufficient to displace the free rotor’s rotation axis from the camera’s principal axis. Therefore, an image stream that is rotation stabilized will appear to orbit about an unknown rotation center. The algorithm was extended and shown to be able to compensate for image eccentricity optimally. Thirdly, the utility of event-based cameras on free rotors was explored. The Contrast Maximisation (CM) framework was identified as a potential solution for simultaneous image and rotation estimation on free rotors. The impact of aggregation functions on the framework was studied and improved upon. Motion blur in event images was also defined, and its impact on rotation estimation was studied. Lastly, as events are spatiotemporal, aggregation of events into an image depends on motion priori for motion compensating events prior to aggregation. This is typically provided for by an inertial sensor or solved for simultaneously as in the CM framework. However, both methods are unsuitable for free rotors due to the tight computation deadline of high dynamic flight. Furthermore, the smoothing of motion-compensated images due to actual event detection time latency in event cameras severely limits the performance of feature-based methods at high angular rates. As such, a direct angular rate estimator is proposed to circumvent the need for motion priors in camera state estimation and sidesteps problematic smoothing of features in the spatial domain due to motion blur.