Propagation of Uncertainty in Anticipatory Image Exploitation Using Polynomial Chaos Random Process Representations

The ability to accurately anticipate target behavior on the basis of surveillance data is critical in many military and civilian contexts. Information regarding target behavior may be drawn from a variety of sources, each of which suffers from uncertainties in the form of noise, inaccuracies, and outright errors. This proposal seeks to develop novel methods for dealing with this uncertainty by vertically integrating uncertainty models in a common framework through all levels of data processing, by adapting uncertainty models over time to incorporate newly observed behaviors and interactions, and by leveraging powerful new adaptive processing techniques. The resulting technology will propagate uncertainties from inputs and models, producing a distribution over anticipated behaviors and a characterization of the most likely future target tracks and associated likelihood measures. These distributions are used to detect objects whose behavior pattern is anomalous relative to the models learned from previous observations. BENEFIT: The ability to accurately anticipate vehicle behavior on the basis of surveillance data is critical in many military and civilian contexts. In particular, behavior anticipation can form the basis of algorithms that identify vehicles whose behavior is atypical or is consistent with some undesirable behavior. In a military setting, a number of behavioral cues are often associated with insurgent operations, including reconnaissance, suicide bombings, and deployment of improvised explosive devices. Similarly, in a civilian setting atypical driving behavior is frequently the initial cause for traffic stops that uncover more serious illicit activities. Examples include unusually slow or excessively fast speeds, erratic driving, failure to stop at traffic lights, etc.