Stochastic Algorithms for Visual Tracking

Following objects as they shift and change shape across video frames is a core challenge in computer vision. In recent years, randomized methods such as particle filters and the Condensation algorithm have markedly advanced what visual tracking can achieve. This book sets out a single framework for tracking with particle filters. It introduces partitioned sampling, a newer technique designed to ease the difficulties that high-dimensional problems pose for standard particle filters.

The author also presents contour likelihoods, a set of models that score object shape, colour and motion using the statistical behaviour of image features. Because these likelihoods are statistical, they fit naturally into stochastic algorithms. A recurring idea throughout is that probability lets the output of these algorithms be read as the computer's "belief" about what is happening in the scene.

Written for students, researchers and practitioners in computer vision, the book requires only a basic understanding of probability theory.