Date: September 24, 2004

Time: 3:00 pm

Location: GCATT Room 325

Speaker(s): Volkan Cevher

Title: Sequential Monte-Carlo Markov Chain Methods for Visual Tracking using a Robust Online Appearance Model

Abstract:
Sequential Monte-Carlo methods, also known as particle filters, are
attractive statistical tools to recursively update a non-Gaussian and
possibly intractable probability density function (pdf) using the Bayes
rule. A continuous state vector pdf is represented by discrete samples
with random support, cleverly constructed so that accurate inferences
can be achieved with moderate computational power. In this talk,
particle filtering solutions on state-spaces are briefly discussed first
with emphasis on proposal strategies. Then, the visual tracking problem
is defined and a quick overview of the existing solutions is given. An
online appearance model is presented that can adapt to the chances in
the target appearance while building a natural stable model for the
target. The model consists of a mixture of stable image structure,
learned over time; a wandering component, based on two frame variations;
and a fixed model, corresponding to a target template. Finally, a
solution to the visual tracking problem is discussed using this
appearance model within the Bayesian framework.

Bio:
Volkan CEVHER was born in Ankara, Turkey, in August 1978. He received his B.S. Degree in Electrical Engineering from Bilkent University, Ankara, Turkey in 1999. He is currently studying towards the Ph.D. Degree in Electrical Engineering in Georgia Institute of Technology under the supervision of James H. McClellan. His current research interests include Monte Carlo Markov chain methods, target tracking models, adaptive filters, time frequency distributions, fractional Fourier transform, and array signal processing.

Date:	September 24, 2004
Time:	3:00 pm
Location:	GCATT Room 325
Speaker(s):	Volkan Cevher
Title:	Sequential Monte-Carlo Markov Chain Methods for Visual Tracking using a Robust Online Appearance Model
Abstract:	Sequential Monte-Carlo methods, also known as particle filters, are attractive statistical tools to recursively update a non-Gaussian and possibly intractable probability density function (pdf) using the Bayes rule. A continuous state vector pdf is represented by discrete samples with random support, cleverly constructed so that accurate inferences can be achieved with moderate computational power. In this talk, particle filtering solutions on state-spaces are briefly discussed first with emphasis on proposal strategies. Then, the visual tracking problem is defined and a quick overview of the existing solutions is given. An online appearance model is presented that can adapt to the chances in the target appearance while building a natural stable model for the target. The model consists of a mixture of stable image structure, learned over time; a wandering component, based on two frame variations; and a fixed model, corresponding to a target template. Finally, a solution to the visual tracking problem is discussed using this appearance model within the Bayesian framework.
Bio:	Volkan CEVHER was born in Ankara, Turkey, in August 1978. He received his B.S. Degree in Electrical Engineering from Bilkent University, Ankara, Turkey in 1999. He is currently studying towards the Ph.D. Degree in Electrical Engineering in Georgia Institute of Technology under the supervision of James H. McClellan. His current research interests include Monte Carlo Markov chain methods, target tracking models, adaptive filters, time frequency distributions, fractional Fourier transform, and array signal processing.