A Comprehensive Analysis of Tracking as a Data Association Problem
Eric Poblenz · eScholarship (California Digital Library) · 2015
Algorithms based on traditional notion of tracking as a state estimation problemyield just a single interpretation of the data. For some applications, the ability toidentify ambiguities and compare different interpretations using a well-defined measureof confidence is critical. Such applications require a direct solution to the data associationproblem in order to characterize the relevant uncertainty. This notion of trackinghas received relatively little attention largely due to a failure to recognize its utilitybeyond maintaining the state estimation process. As a result, the options available tothe practitioner are limited and the performance of statistical data association modelsis not well understood, especially in terms of the quality of the sample they produce.This work has sought to change that by developing a new data associationmodel that extends the scope and flexibility of existing models. The questions of howto specify an objective prior distribution over data association hypotheses and howto efficiently perform inference on the high-dimensional posterior distribution are verymuch open. To help provide answers, we considered numerous different priors, includingBayesian nonparametric models and several models never before applied to tracking.With regard to inference, we considered various implementations of Markov chain MonteCarlo (MCMC) and population Monte Carlo (PMC) samplers. A comprehensive evaluationwas performed in the context of a wide-area radar surveillance application.