Authors : B. Wohlberg (Los Alamos National Laboratory, USA), P. Rodriguez (Pontificia Universidad Catolica del Peru, Peru)
Chapter Description
Robust PCA has been very successful for background modeling in video from a stationary camera, the underlying model being that the background occupies a low-dimensional subspace, while the foreground is represented by sparse deviations from this subspace. These methods do not perform well, however, in
the presence of a non-stationary background resulting from a moving camera since in this case it is more accurate to consider the background as dynamically traversing a low-dimensional but non-linear manifold. A promising approach to modeling this geometry is by estimating local tangent subspaces to this manifold via endogenous sparse representations, based on the same ideas that inspired the state-of-the-art sparse subspace clustering technique. Computational experiments indicate that this approach provides substantially better performance than RPCA in modeling video with a moving background.