AN APPROXIMATE EXPECTATION - MAXIMIZATION LIKE APPROACH TO SPATIO - TEMPORAL BELIEF - PROPAGATION FOR MOVING OBJECT DETECTION

Open Access
Author:
Rengarajan, Bharath
Graduate Program:
Electrical Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
June 16, 2011
Committee Members:
  • Dr Robert Collins, Thesis Advisor
  • Dr David Miller, Thesis Advisor
Keywords:
  • Moving Object Detection
  • Markov Random Field
  • Belief Propagation
  • Expectation Maximization
Abstract:
Graphical models offer a convenient framework to model vision problems like motion estimation. The idea of modeling a pixel’s motion likelihood using Belief Propagation (BP) in a 3D Markov Random Field (MRF) has been previously explored by applying a message passing algorithm in a 6-connected spatio-temporal neighborhood. This thesis extends that framework by considering improvements to the inference algorithm. In particular, we extend the results of the discrete-BP approach adopted previously to continuous-BP with parametrized compatibility functions, and re-formulate the problem to jointly perform both approximate inference and parameter estimation within an approximate Expectation-Maximization (EM) framework. The accuracy and efficiency of the proposed approach is validated on simulated data and on various real-world video sequences. Comparative results are presented for the discrete-BP and the EM-BP approaches.