VASC Seminar Announcement
            =========================

Date: Friday, 9/17/99
Time: 1:30-2:30
Place: Smith Hall 2nd Floor Common Area

Speaker:  Jim Rehg
          Compaq Corporation
          http://www.crl.research.digital.com/who/people/rehg/bio.htm


Title: Progress in Figure Tracking and Dynamics Learning

Abstract:

In this talk I will describe two new results in visual tracking and motion
analysis for complex objects such as the human figure.  

A major limitation in current figure tracking systems is the lack of an
adequate dynamical model.  Since analytic models are either too simplistic
(i.e. constant velocity) or too cumbersome (i.e. biomechanical), we address
learning models from motion data.  Switching linear dynamic system (SLDS)
models are attractive because complex time-varying dynamics can be
constructed from simple linear building blocks.  Learning of SLDS models
follows the standard EM approach.  Unfortunately, exact inference (E-step)
in SLDS is intractable.  I will describe two approximate inference
algorithms that make learning feasible.  This is joint work with Vladimir
Pavlovic.

A second barrier to robust tracking is track initialization and failure
recovery.  I will describe a novel, efficient registration algorithm based
on dynamic feature ordering.  This algorithm can exploit the structure of
complex object models to dramatically lower the cost of registration.  It is
particularly suited to track initialization, where prior information is
weak, and track recovery, where measurement data is ambiguous.  This is
joint work with Tat-Jen Cham.

Jim Rehg received his PhD from CMU in 1995, studying under Prof. Takeo
Kanade.  He is at the Cambridge Research Lab of Compaq Computer Corporation,
where he leads the Vision-Based Human Sensing project.