Modeling and recognizing situations of interest in surveillance applications

Today's surveillance systems are very powerful in performing the process of object assessment, i.e., to estimate an object's position and attributes over time. However, the interpretation of the object's behavior, i.e., the situation assessment process, is still done by human experts. In this article, we describe an approach of how expert knowledge about situations of interest can be modeled in a situational dependency network (SDN). Based on the SDN, we present an approach of constructing a probabilistic model, namely a dynamic Bayesian network (DBN). We will describe in detail how the structure and the parameters of such a DBN can be specified automatically. The DBN can then be applied to observations made over time. Finally, we will show some evaluation results on simulated observation data with different amount of noise and show that the model yields the expected results.