| dc.description.abstract | The ubiquitous nature of CCTV Surveillance cameras means substantial amounts of
data being generated. In case of an investigation, this data must be manually browsed
and analysed in search of relevant information for the case. As an example, it took
more than 450 detectives to examine the hundreds of thousands of hours of videos in
the investigation of the 2011 London Riots: one of the largest the London's MET police
has ever seen. Anything that can help the security forces save resources in investigations
such as this, is valuable. Consequently, automatic analysis of surveillance scenes is a
growing research area.
One of the research fronts tackling this issue, is the semantic understanding of the scene.
In this, the output of computer vision algorithms is fed into Semantic Frameworks, which
combine all the information from different sources and try to reach a better knowledge of
the scene. However, representing and reasoning with imprecise and uncertain information
remains an outstanding issue in current implementations.
The Demspter-Sha er (DS) Theory of Evidence has been proposed as a way to deal with
imprecise and uncertain information. In this thesis we use it for the main contributions.
In our rst contribution, we propose the use of the DS theory and its Transferable Belief
Model (TBM) realisation as a way to combine Bayesian priors, using the subjectivist
view of the Bayes' Theorem, where the probabilities are beliefs. We rst compute the
a priori probabilities of all the pair of events in the model. Then a global potential is
created for each event using the TBM. This global potential will encode all the prior
knowledge for that particular concept. This has the bene t that when this potential is
included in a knowledge base because it has been learned, all the knowledge it entails
comes with it. We also propose a semantic web reasoner based on the TBM. This reasoner consists of an
ontology to model any domain knowledge using the TBM constructs of Potentials, Focal
Elements, and Con gurations. The reasoner also consists of the implementations of the
TBM operations in a semantic web framework. The goal is that after the model has been
created, the TBM operations can be applied and the knowledge combined and queried.
These operations are computationally complex, so we also propose parallel heuristics to
the TBM operations. This allows us to apply this paradigm on problems of thousands
of records.
The nal contribution, is the use of the TBM semantic framework with the method to
combine the prior knowledge to detect riots on CCTV footage from the 2011 London
riots. We use around a million and a half manually annotated frames with 6 di erent
concepts related to the riot detection task, train the system, and infer the presence of riots
in the test dataset. Tests show that the system yields a high recall, but a low precision,
meaning that there are a lot of false positives. We also show that the framework scales
well as more compute power becomes available. | en_US |
