Tracking and Identifying of Multiple Targets

Abstract

There are many statistical methods of tracking single and multiple targets; this manuscript will focus on the state estimation problem. Ideally, a generalization of the recursive Bayes non-linear filter would track and resolve the state(s) of single or multiple targets, but that is currently computationally intractable. The Probability Hypothesis Density (PHD) makes the tracking problem computationally feasible by propagating only the first-order multi-target statistical moments by using a particle filter implementation for the PHD. The problem then becomes one of estimating the targets’ state based on the output of the PHD when using a particle filter implementation.

This paper describes one heuristic method for obtaining a state estimator from the PHD. The approach used in this paper, based on the Expectation-Maximization (EM) algorithm, views the PHD distribution as a mixture distribution, and the particles as an i.i.d. sampling from the mixture distribution. Using this, a maximum likelihood estimator for the parameters of the distribution can be generated. The EM seems to work fairly well, particularly when targets are well spaced.