| |
Abstract: Since many years,
motion estimation (ME) has been investigated in two major
domains: video compression and video analysis. In video
compression, motion estimation is aimed at reducing the
temporal redundancy in the successive frames of a video
stream. This temporal redundancy can be represented by the
appearance of the same pattern, or object, in several
successive frames. Thus, in order to reduce the computation
time for spatial compression, the spatial compression of a
pattern is done once and it is only the "estimated" motion
vector of this pattern which is transmitted in the next
frames.
In video standards like MPEG4, an object approach of the
compression has been planified. A video scene is thus
decomposed in fast moving objects and "sprites" or almost
static objects and background. The evolution of video
standards, based on the old (MPEG2) block-based compression
(block-matching), but enhanced with a relaxation around the
"block splitting" constraint, has leaded to very efficient,
"non-object oriented" compression standards like H264.
Within an object-oriented (OO) compression framework, first,
we have investigated the possibility to estimate the motion,
not only of the DCT blocks (or macroblocks), but of segmented
objects.
In our OO approach of the ME, a segmentation of objects of
interest is done by means of a Potts-Markov modeling and
Bayesian estimation. In order to improve the computation
speed, the segmentation is performed in the orthogonal wavelet
domain, and the Potts model is tuned to the wavelet subbands
orientations. A spatial correlation between successive frames
enables also to increase the segmentation speed; this is
simply realized by initializing the segmentation of each frame
by the segmentation result of the former frame. At this point
our motion estimation stands as "object" or "region"-matching.
Naturally (or inherently), the multiresolution of a wavelet
transform provides us with a tool that is perfectly adapted to
a hierarchical analysis of the frame and of the ME. We recall
here the advantages of a hierarchical approach in ME :
- A progressive, thus fast, transmission of motion vectors
(MV), and of frames, at low scales
- the robustness of the hierarchy where motions at a scale are
highly related to motions at a coarser scale
- The possibility to limit the number of acknowledgements in
video bitstreams transmission, due to the inherent robustness
of estimated motion vectors.
But the orthogonal wavelet transform, practically, is limited
by:
- its dependency to translation
- its limitation in resolution at low (coarse) scales
- its non-capacity to analysis
Here we introduce, in our investigation of the motion
estimation, wavelet families tuned to the analysis of motions,
and in particular to speed.
The continuous wavelet transform (CWT) provides us with the
same resolution of the filtered frames at all scales. It also
has the ability to detect objects at specific speeds and
rotations, in addition to the traditional translation and
scale parameters. These so-called spatio-temporal wavelets
have been mainly developed on the basis of the Morlet wavelet
and are being studied with other filters like the conical and
Cauchy wavelets. These kernels provide more anisotropic
wavelets which is here the quality required to detect
"singularities" or specific parameters in oursince years video
scenes. Brief
Biography of the speaker:
Dr. Patrice Brault graduated from the Conservatoire
National des Arts et Metiers, CNAM, in Electrical Engineering.
Before joining the Centre National de la Recherche
Scientifique, CNRS, in 1998, he has been working in the
telecommunications area, mainly for Matra Communications,
Apple Europe Research, and the Laboratoire d’Electronique
Philips, LEP, where he participated to the development of the
first complete MPEG2 digital television broadcast system.
His main research interests are signal and image processing,
and, in particular, fractals, wavelets, and Bayesian methods
applied to shape recognition, motion estimation, segmentation,
and video compression. He owns a PhD in Physics, with a
speciality in Electronic Systems and Information Theory, and
is presently at the Laboratoire d’Electronique Fondamentale,
IEF, University of Orsay Paris-sud, CNRS UMR8622, FRANCE. IEEE
and SIAM member. |
