Taking Advantage of Group Behavior When Tracking Multiple Threats in Cluttered Surveillance Data

Abstract

Threats are composed of some process or plan being carried out by a group of people with an end goal that is generally to cause harm. Some examples of these kinds of threats are terrorist attacks, military actions, or stock fraud. These threats can be modeled stochastically with help from experts within the relevant field. We model these threats with a hypothesis as to how these events will unfold along with a method for observing the unfolding threat. We will use this model to detect the threat before its completion and, ideally, allow for preemptive action against it. The models used for threats in this paper are variations of Hidden Markov Models (HMMs) with sparse observation emission (rare as compared to the expected process length). The population observed is assumed to be organized into groups called “cliques”. Rather than tracking an individual's involvement probability as was done in a related effort, we track a clique's (group's) involvement probability across all threats using a Bayesian update equation and conditioning on association events between the observations and the set of measurement generating HMMs (threat and clutter processes). We assign an individual's involvement probability conditionally based on that of their group, and thence the state of each threat process then its state is estimated using a bank of Bernoulli filters. This will allow us to accurately detect multiple threat processes within a single stream of observations (most of which will be clutter).