Abstract
The ability to localize, i.e., determine the position and orientation of a Unmanned Aerial Vehicle (UAV) with respect to a known frame of reference, is a basic requirement for autonomous flight. Common solutions for providing a UAV with localization ability have relied on the availability of an infrastructure built for this purpose, usually based on an arrangement of radio emitters, predominantly Global Navigation Satellite Systems (GNSSs). However, disruptions in the radio signal path, as well as actions taken by an adversary, such as spoofing and jamming, may hinder localization accuracy. This thesis focuses on UAV localization, in environments lacking infrastructure for that purpose, specifically using a low-size, weight and power (SWaP) sensor system consisting of a camera, an Inertial Measurement Unit (IMU), and a magnetometer. The challenges limiting this approach are associated with the difficulty of relating UAV environment measurements to a map, due to not only differences between the appearance of the map representation and the environment as observed using onboard sensors, but also natural ambiguities such as perceptual aliasing. This thesis addresses three specific problem areas and demonstrates a full localization solution running in real time on a small UAV. First, the thesis addresses the problem of how to perform localization with respect to an orthophoto map using a camera whose orientation is not strictly vertical. A method is presented for allowing variation in camera view direction by orthoprojecting camera images to a top-down view based on a planar assumption of the ground under the UAV. This would be an adequate assumption when flying over relatively flat terrain, as demonstrated through experimentation on real data. Second, this thesis addresses the problem of seasonal appearance change, where we learn a function for assessing the correspondence between an image acquired by an UAV and an orthophoto map by proposing a method that is tolerant to seasonal appearance change in the operating environment. The proposed method exceeds the state-of-the-art in the literature both in terms of the time to convergence and localization error. Third, this work addresses the wake-up robot problem. For this purpose, an approach is presented for learning a model to extract a compact descriptor vector representation from both a UAV image and from a map, thus enabling very fast confirmation or rejection of pose hypotheses, which allows localization to occur over large areas without knowledge of the initial pose. The presented method alleviates the computational challenges inherent in the problem of localization over a large area with an unknown prior starting position and orientation. The method also enables operation of a small UAV on a map covering an area of 100 square kilometers without requiring knowledge of the initial pose while tolerating seasonal appearance change and resolving ambiguities due to perceptual aliasing. Finally, the operation of the algorithm developed for the wake-up robot problem running on a small UAV is demonstrated in real time using real data. The thesis concludes by characterizing a number of open issues related to the problem domain.
Translated title of the contribution | Miehittämättömien lentävien laitteiden infrastruktuuririippumaton paikannus |
---|---|
Original language | English |
Qualification | Doctor's degree |
Awarding Institution |
|
Supervisors/Advisors |
|
Publisher | |
Print ISBNs | 978-952-64-1946-6 |
Electronic ISBNs | 978-952-64-1947-3 |
Publication status | Published - 2024 |
MoE publication type | G4 Doctoral dissertation (monograph) |
Keywords
- localization
- unmanned aerial vehicle
- robotics