Hybrid Correlational Graphical Models for Reasoning in Detecting Systems

Abstract

Using probabilistic graphical models to deal with uncertainties by modeling relationships among detecting objects is a common method for event detecting systems. However, not all relations are captured accurately by former graphical models. This paper presents a hybrid correlational model for typical abnormal event detecting systems that have correlated objects. It captures the OR relation of multiple influences from different sources of the abnormal event. An algorithm based on message passing is developed for efficient reasoning in the model. Analysis and experiments are provided to compare it with former graphical modeling by results on the detecting objects that lack of local evidence, and by their sensitivity to the occurrence of abnormal event.