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.