MultipleMotion Analysis Using
3D Orientation Steerable Filters
In this paper, we study the characterization of multiple motions from the standpoint of
orientation in spatiotemporal volume. Using the fact that multiple motions are equivalent
to multiple planes in the derivative space or in the spectral domain, we apply a
new 3D steerable filter for motion estimation. This new method is based on the decomposition
of the sphere with a set of overlapping basis filters in the feature space. It
is superior to principal axis analysis based approaches and current 3D steerability approaches
in achieving higher orientation resolution. Our approach is more efficient and
robust than a similar spatiotemporal Hough transform and outperforms existing EM
algorithms applied in the derivative space.
In occlusion estimation, we use an eigenvalue analysis based multi-window strategy to
detect and to eliminate outliers in the derivative space. This technique purifies input
data and improves therefore the precision of the estimation results. Furthermore, based
on the spatial coherence in image sequences we use the “shift-and-subtract” technique
to localize occlusion boundaries and to track their movement in occlusion sequences.
Our technique can be also used to distinguish occlusion from transparency and to decompose
transparency scenes into multi-layers.