Abstracts for 2003 WOMOT Tutorials

June 21, 2003

Time

Tutorial

8:00 - 9:00 a.m.

"Fundamentals of Small Target Tracking"

T. Kirubarajan
McMaster University
Hamilton, Ontario, CANADA
 

Target tracking consists of data association, filtering and fusion, and a number of algorithms are available to handle these issues in tracking small (or point) targets. In this presentation, we will review different algorithms, including the Kalman filter, Interacting Multiple Model (IMM) estimator, Probabilistic Data Association (PDA) algorithm, Multiple Hypothesis Tracking (MHT) algorithm and assignment techniques. Possible fusion architectures and performance measures for target tracking will be discussed as well. Simulation results on some realistic tracking scenarios will be presented.
9:00 - 9:55 a.m.

"Engineering Statistics for Multi-Object Tracking"

Ronald P. Mahler
Lockheed Martin Tactical Systems
Eagan, MN, USA


Progress in single-sensor, single-target problems has been greatly aided by the existence of a systematic, rigorous, and yet practical engineering statistics.  One might expect that the same would be true for multisensor-multitarget problems.  Surprisingly, this has not been the case, even though a comprehensive statistical foundation for multi-object problems—point process theory—has been in existence for decades.   The primary purpose of this tutorial is to provide a brief, high-level overview of finite-set statistics (FISST), the ''engineering friendly'' version of point process theory that Dr. Mahler introduced in 1994.  FISST is engineering-friendly in that it is geometric, and preserves the “Statistics 101” formalism that signal processing engineers already understand.  Its core is a multisource-multitarget differential and integral calculus, based on the fact that belief-mass functions are the rigorous multisensor-multitarget counterparts of probability-mass functions.  One novel consequence is that FISST encompasses expert-system approaches such as fuzzy logic, the Dempster-Shafer theory, and rule-based inference.  A secondary purpose of the tutorial is to demonstrate the relevance of FISST to practical applications such as robust INTELL multisource NCTI, multitarget tracking, and performance evaluation.  A third purpose is to address such few criticisms of FISST as there have been.  The optimality and simplicity of Bayesian methods can be taken for granted only within the confines of standard applications addressed by standard textbooks.  This tutorial will show that when one ventures out of these confines—especially in multitarget problems—complacency can lead to serious problems.

9:55 - 10:05 a.m. Break
10:05 - 11:00 a.m

"Sequential Monte Carlo Methods for Multi-Object Tracking"

Simon Maskell
QinetiQ and Cambridge University
ENGLAND

Sequential Monte Carlo methods, or particle filters, provide a powerful Bayesian methodology for sequential inference in nonlinear non-Gaussian state-space systems. After an introduction to the approach, in an attempt to improve intuition as to how particle filters can be used to track multiple targets, two thrusts of current research will be described within the context of importance sampling. Firstly, the crucial role of the choice of importance distribution will be described in terms of changing the memory of a system that the samples must populate. Secondly, the use of analytic integration to reduce the Monte-Carlo variance will be explained through consideration of the resultant reduction in dimensionality of the state-space that the samples inhabit.

11:00 - noon

"Visual Tracking of Multiple Objects Using Particle Filters"

John MacCormick
HP Labs
Palo Alto, CA, USA

After reviewing the fundamental theory of multi-object particle filters, we'll describe the popular approaches to using particle filters in visual tracking problems. In particular, we will investigate the answers to the following two questions: (i) How can all appropriate data from a video sequence be incorporated into the inference process, so that particles hypothesizing different numbers of objects can be treated coherently? (ii) How can object births and deaths be modeled without either destroying the probabilistic validity of the particle filter or creating an excessive computational burden? In addition, we will discuss strategies for variance reduction that are particularly suited to visual tracking problems.