A Distributed Video Surveillance System to Track Persons in Camera Networks

This thesis focuses on the topics of information visualization in a video surveillance system and on distributed person re-identification. Visualizing the proper information in a concise and informative fashion is a very challenging tasks that is mainly driven by the situation, the task that has to be performed and last but not least, by the operator that is using the system. Designing a successful system that is able to support all of these requirements and constraints is the first goal of this thesis. Towards this end we design and develop four system prototypes and evaluate each of them by means of standard Human-Computer Interaction principles. We show that this approach leads to an advanced system that is capable to support the task of tracking a person moving through multiple cameras field-of-views using only a single display. The advanced visualization system was built to support the task of tracking a person through multiple overlapping cameras. However, in a real scenario this is not always feasible and we have to deal with disjoint cameras, hence, the system may fail to track the same person moving across them. In light of this, we propose three different methods to tackle the person re-identification problem so as we can re-associate a person that moves out from one camera and then reappears in another one at a different time instant. The first method builds a discriminative signature for each person that is matched by using a robust distance measure. The second method studies the transformation of features across cameras, while the last one builds upon the idea that as features get transformed so is the distance between them. Finally, we consider the issues of a fully centralized camera-camera re-identification system and introduce a distributed re-identification framework. For each approach, experimental results on public benchmark datasets are given.