|
|
Foundations of Robotics Seminar, March 1, 2011
Time
and Place | Seminar Abstract
Optimization Methods for Nonlinear Data Association
Matthew Travers
LIMS Lab
Northwestern University
Tuesday, March 1, 2011
NSH 3305
Talk 4:30 pm
This talk focuses on nonlinear data association. Data association is a field which addresses the uncertainty in the origin of measurements due to complex environments. Over the past fifty or so years, considerable attention has been given to data association for multi-body systems. Popular techniques, such as the multiple hypothesis tracker or Markov-chain Monte-Carlo data association, consider discrete-time systems and focus on solving an underlying combinatorial optimization. The problem with these approaches is computational intractability as the number of measurements gets large. The subject of this talk is an algorithm which recasts the multi-body data association problem as a continuous optimization in order to take advantage of the maximum principle. The purpose of doing so is to determine an efficient method to search the space of possible solutions, which grows exponentially with the addition of new measurements. The main results to be presented are first- and second-order adjoint formulations which reduce the multi-body data association problem (associating measurements to their object of origin) down to a fixed number of integrations, regardless of the number of measurements.
Matthew Travers is a PhD candidate at Northwestern University. He completed his BS in Engineering Physics and MS in Electrical and Computer Engineering at the University of Colorado at Boulder. Matt's current research interests include optimization techniques and their role in optimal control as well as optimal estimation.
The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.