Abstract

Aggression and violence can be used in a very broad way in our daily life. If you pick up a newspaper or watch the evening news, there is always some instance of aggression. The impact of any form of aggression can vary from simple pain to great depression. To cope with aggression detection in an automated way there are some technical challenges that can vary from human and group tracking, speech/audio recognition, facial recognition to behavior recognition. The main challenge for this thesis is to put all the objects and concepts in our domain in the right context of each other. A first step is to model the train environment domain. To cope with the task of detecting aggression in trains the Dutch Railways (Nederlandse Spoorwegen, NS) and the Man-Machine Interaction Group at the Delft University of Technology are working together on this project. The main challenge for this thesis is to put the main objects and concepts in our domain in the right context of each other. An information system that is capable of detecting aggression must first be aware of the objects in the environment, their capabilities, and their relationships and perform automated semantic interpretation of events in the environment. This can be done by designing an aggression ontology, which is the first task. Ontologies are computer-stored specifications of concepts, properties, and relationships that are important for describing an area of expertise(domain).
The next task is to design a Bayesian Network(BN). The choice for a Bayesian approach was such that we had to find an appropriate tool able to represent various sources of uncertain information that describe our problem, and to join them into an inference system. Our proposed Bayesian network will handle incoming data from an annotating process using the concepts from the ontology. The output will give a probability for the 5 aggression scenario classification (neutral, damage, annoyance, danger and sickness). A case study on how real data(acquired from actual footage shot in a train with actors), can be modeled in our Bayesian Network and inferred with will be described and evaluated in this thesis.