Contextual Cues for Causal Visual Tracking

As a consequence of the ever increasing automation, many application domains - such as autonomous driving or visual surveillance - have to deal with vast amounts of visual data. Efficient data processing and subsequent reasoning about the ongoing events requires automated video analysis. An essential requirement for such automated analysis is to accurately localize objects and reliably estimate their trajectories over time, in order to deduce which (inter-)actions are observed by a camera. To address these tasks, numerous visual object tracking paradigms have been investigated over the past few decades. The majority of these approaches, however, focuses only on the dynamics and visual representation of the target itself, neglecting the information gain provided by other contextual cues which are readily available from the recorded visual data.
In this thesis, we investigate the potential of auxiliary scene information, i.e. context, to robustify visual object tracking. To this end, we exploit often neglected information sources to build intuitive, yet very accurate and efficient tracking models. These models cover both appearance-based and geometric context to address several limitations of existing work. Appearance, on the one hand, can be used to reduce the risk of drifting in the case of visually ambiguous scenarios. Leveraging geometric prior knowledge and observed scene dynamics, on the other hand, allows to model plausible movements of missed or otherwise undetected objects which can be exploited to resolve occlusions. We rely on these context cues to build causal visual object trackers, which are suitable for time-critical applications. To demonstrate both the benefits and limitations of each context-aware model, we conduct detailed evaluations on challenging real-world test scenarios.