PCA is a classical method to learn common information of a sequence of images. Especially in video analysis, this requires the computation of the (thin) SVD of large matrices. We present a method, based on an efficient SVD computation for augmented matrices, that allows to adapt the number of relevant singular vectors while analyzing a sequence of images. This technique allows the detection of moving elements in surveillance videos with varying background.

This presentation is part of Minisymposium “MS2 - Interpolation and Approximation Methods in Imaging (4 parts)”
organized by: Alessandra De Rossi (University of Torino) , Costanza Conti (University of Firenze) , Francesco Dell'Accio (University of Calabria) .

Authors:

Tomas Sauer (University of Passau)

Keywords:

computer vision, machine learning, numerical linear algebra