Ecole Centrale de Paris
6-8 avenue Blaise Pascal
Cite Descartes, Champs-sur-Marne,
77455 - Marne-La-Vallee, FRANCE
Modeling and Subtraction of Dynamic Scenes
Modeling Dynamic Scenes is an important component of many vision systems. In this talk, we propose two different concepts to address such a demand. First, we proposed a predictive mechanism that is based on an auto-regressive model to capture and predict the behavior of such scenes.
The central idea is to treat the input signal as a time series where previous observations are used to predict the new observation that is then compared with the actual one towards detection of objects that do not belong to the scene. Second, we consider the use of optical flow to model the dynamic characteristics of the scene according to probability density function.
Inherent ambiguities in the computation of features are addressed by using a data-dependent bandwidth for density estimation using kernels. Given a new observation, optical flow is determined and along with visual appearance is used to determine the probability for a given pixel to be part of the dynamic scene. Promising results demonstrate the potentials of these such approaches. Joint Work with Anurag Mittal, Antoine Monnet and Visvanathan Ramesh.
Wednesday, October 13, 2004
DCC 324 - 4:00 p.m. to 5:30 p.m.
Refreshments at 3:30 p.m.