2.5D Visual Tracking and Layered Image Analysis


Research Overview

There exists a theoretical gap between model-based tracking of complex objects and early motion estimation, namely, the selection and initialization of models. It seems clear that some form of early analysis should help one select and initialize models, but this remains a largely unexplored topic. Allan Jepson, Michael Black and I have been working on methods for extracting effective representations of visual motion that provide an efficient characterization of the principal moving components of a scene and their relative depths. In future research we hope to use this representation to infer the occurrence of plausible models for subsequent model selection and refinement.

To provide stable image descriptions through time, Allan Jepson, Thomas El-Maraghi (University of Toronto) and I have developed an approach to learning 2D models of image appearance. The approach works with an online version of the EM algorithm to identify stable image structure during tracking. In this way the most stable regions can play the most significant role in the tracking, to facilitate tracking over long image sequences with precise image alignment. This work was awarded Runner-Up for the Best Paper at the IEEE Conference on Compuer Vision and Pattern Recognition in 2001.


Related Publications


Return to David Fleet's home page.