A Bayesian Nonparametric Approach to Modeling Motion Patterns

The most difficult—and often most essential—aspect of many interception and tracking tasks is constructingmotion models of the targets to be found. Experts canoften provide only partial information, and fitting parametersfor complex motion patterns can require large amountsof training data. Specifying how to parameterize complexmotion patterns is in itself a difficult task.In contrast, nonparametric models are very flexible andgeneralize well with relatively little training data. We proposemodeling target motion patterns as a mixture of Gaussianprocesses (GP) with a Dirichlet process (DP) prior overmixture weights. The GP provides a flexible representationfor each individual motion pattern, while the DP assigns observedtrajectories to particular motion patterns. Both automaticallyadjust the complexity of the motion model basedon the available data. Our approach outperforms several parametricmodels on a helicopter-based car-tracking task ondata collected from the greater Boston area.