The Conditionalizing Identity Management Bayesian Filter (CIMBal)

Abstract

We present a large-scale data association tracker that can handle variable numbers of world objects and measurements. Large-scale data association problems arise in surveillance, wildlife monitoring, and applications of sensor networks. Several approaches have recently been proposed that represent the uncertainty in data association using a parameterized family of distributions on the set of permutations. Whereas these approaches were restricted to a fixed and known number of objects (and sometimes measurements), we generalize these approaches to varying numbers of objects and measurements. We also present a modification that allows one to focus on a set of objects of interest, while maintaining data association with all other objects that may be confused with these objects of interest. We justify the approach with an analysis and show experiments on a large-scale simulated tracking sequence.