Abstract:

There has been growing interest in algorithms capable of learning models from large volumes of multidimensional data, using statistical, geometrical and dynamical information. There are many domains of application for such algorithms, e. g. in exploratory data analysis, computer vision, system identification, control, computer graphics and multimedia databases.

While the linear case can be solved by the well-known Principal Component Analysis technique, the non-linear case is more complex. Recently, there have been advances in algorithms that approximate the data through manifold learning. The present works fits this frameworks, with emphasis on the problem of motion tracking – particularly in video sequences – assuming that the whole observation space is not occupied, but rather a manifold embedded in that space.

This thesis proposes a manifold learning algorithm, named Gaussian Process Tangent Bundle Approximation (GP-TBA). This algorithm can deal with arbitrary manifold topology by decomposing the manifold into multiple local models, while also providing a probabilistic description of the data based on Gaussian process regression.

The model provided by GP-TBA is also used to simplify the motion tracking problem, for which a multiple filter architecture, using e. g. Kalman or particle filtering, is described. The GP-TBA algorithm and the filter bank framework are illustrated with experimental results using real video sequences.

Date: 2006-Nov-16     Time: 16:00:00     Room: 336

For more information:

• las@inesc-id.pt