Universität Tübingen
Leveraging Metadata for Computer Vision on Unmanned Aerial Vehicles
Abstract
The integration of computer vision technology into Unmanned Aerial Vehicles (UAVs) has become increasingly crucial in various aerial vision-based applications. Despite the great significant success of generic computer vision methods, a considerable performance drop is observed when applied to the UAV domain. This is due to large variations in imaging conditions, such as varying altitudes, dynamically changing viewing angles, and varying capture times resulting in vast changes in lighting conditions. Furthermore, the need for real-time algorithms and the hardware constraints pose specific problems that require special attention in the development of computer vision algorithms for UAVs. In this dissertation, we demonstrate that domain knowledge in the form of meta data is a valuable source of information and thus propose domain-aware computer vision methods by using freely accessible sensor data. The pipeline for computer vision systems on UAVs is discussed, from data mission planning, data acquisition, labeling and curation, to the construction of publicly available benchmarks and leaderboards and the establishment of a wide range of baseline algorithms. Throughout, the focus is on a holistic view of the problems and opportunities in UAV-based computer vision, and the aim is to bridge the gap between purely software-based computer vision algorithms and environmentally aware robotic platforms. The results demonstrate that incorporating meta data obtained from onboard sensors, such as GPS, barometers, and inertial measurement units, can significantly improve the robustness and interpretability of computer vision models in the UAV domain. This leads to more trustworthy models that can overcome challenges such as domain bias, altitude variance, synthetic data inefficiency, and enhance perception through environmental awareness in temporal scenarios, such as video object detection, tracking and video anomaly detection. The proposed methods and benchmarks provide a foundation for future research in this area, and the results suggest promising directions for developing environmentally aware robotic platforms. Overall, this work highlights the potential of combining computer vision and robotics to tackle real-world challenges and opens up new avenues for interdisciplinary research.
Author and committee
dc:creator, dc:contributor.*- Author
-
- Kiefer, Benjamin
Identifiers
dc:identifier.*- Identifier
- hdl:10900/148618