Stochastic Systems Group  

Dynamic Dependency Tests for Object Interaction Analysis
Michael R. Siracusa
SSG, MIT
Consider multiple objects moving in an environment. Given noisy measurements of their position, we wish to determine which, if any, of these objects are interacting and how this interaction evolves over time. We formulate this problem as inference on a class of dynamical models in which interaction is described by changing dependency structure. Specifically, we use a hidden factorization Markov model (HFactMM). This model allows us to take advantage of both structural and parametric changes associated with changes in the state of interaction of a set of objects. We show how inference and learning on an HFactMM can be used in an approximate inference procedure on a more expressive switching linear dynamic system model. Some synthetic examples are presented along with early empirical results on sports data.
Problems with this site should be emailed to jonesb@mit.edu