Appearance Models for Visual Tracking*  

One of the main factors that limits the performance of visual tracking algorithms in computer systems is the lack of suitable appearance models of objects being tracked. The appearance of objects often changes significantly over time. In this work we posit that it is particularly useful to identity the properties of image appearance that remain stable through time, thereby comprising the most reliable properties for visual tracking.

We propose a mathematical framework for learning robust, adaptive, appearance models for complex natural objects. The appearance model is a probabilistic mixture of image structure with different degrees of temporal stability. An on-line EM-algorithm is used to adapt the model parameters over time. An implementation of this approach, based on responses of a steerable pyramid (much like V1 simple cells), will be described and several demonstrations of its performance on image sequences will be given.

*Joint work with Allan D Jepson and Thomas F El-Maraghi.

Dr. David J. Fleet
Department of Computer Science
University of Toronto