Defining dynamic Bayesian networks for probabilistic situation assessment

In surveillance systems, the situation awareness of decision makers is often a crucial point in making appropriate decisions. For supporting the situation assessment process, modules performing an automatic interpretation of the observed environment can be used. However, there is still a need for an optimal solution for the definition of such modules. In this article we describe how situations of interest can be modeled in a human-understandable way and how their existence can be inferred from sensor observations by the use of dynamic Bayesian networks. A crucial point of modeling such networks is the definition of the parameters, namely the conditional probabilities. We present a method for an automatic definition of the parameters that can be easily used by a human operator when designing a new network. By using this approach, we define two example networks that are able to recognize situations of interest in the VIRAT dataset. Finally, the two networks are applied to the VIRAT dataset and we present an evaluation of the performance of the automatic situation assessment.