Thesis-2007-Bhaskar.pdf (23.95 MB)
An integrated block based motion estimation framework for video applications
thesis
posted on 2011-02-17, 10:01 authored by Harish BhaskarMotion Estimation is a popular technique for computing the displacement vectors of objects or attributes between image frames at different time stamps. Motion estimation is critical and forms an integral part of many application domains such as video coding, compression, object
tracking, video indexing, video stabilization, etc. However, the
available motion models are restricted in their generality and have
been tailored for use in specific application areas. The purpose of
this thesis is to propose an integrated block-based motion estimation
framework that serves different real-time video applications including;
object tracking, video stabilization and low level video indexing.
In this thesis, the proposed framework for motion estimation is based
on block matching, a well known strategy for motion estimation particularly
in the domain of video coding and compression. Traditional
block matching techniques are limited in the following ways:
i) block partitioning methods, whether fixed or variable sized, divide
image frames into blocks blindly neglecting the features involved
within.
ii) block search schemes assume restricted translational displacernents
(usually bounded by a search window: time complexity
rises exponentially against search size). iii) during block matching, blocks undergo negligible or null rotntional
motion and cannot suffer complex deformation characteristics
during motion. iv) finally, block matching methods arc incapable of handling occlusion.
The thesis will propose an Integrated block-based motion estimation
framework that handles the aforementioned limitations of existing
schemes. First, we propose a vector quantization based block partitioning
methodology that will extend the quad-tree mechanism, but,
place partitions such that similar or near-similar attributes are clustered
within the same block. In this way we preserve the advantages
of variable block matching by using a well defined quad-tree data
structure and separate regions of interest from the rest of the image
at the same time. At the second level of abstraction, we propose a genetic
algorithm based block search scheine. A genetic algorithm based
search mechanism will present a similar range of computation time irrespective
of the amount of displacement. This will allow the search
space to remain unrestricted and maintain tolerable time complexity.
An immediate extension to the basic framework is presented as a rotation
invariant scheme that is further generalized into a deformation
handling mechanism for motion estimation. We use an affine based
integrated model along-side the genetic algorithm search to match
blocks (in turn image attributes) that may or may not undergo rotation
or deformation during motion estimation. We also present a novel extension to the 2D affine genetic algorithm combination for handling specific 3D rotational changes to blocks. In this thesis, we also propose to integrate a novel motion correction mechanism based
on probabilistic motion modeling for occlusion handling. Finally, for
optimization purposes we combine scale space based architecture to
the framework. This optimization procedure will allow the system to
automatically choose, based on performance metrics, the operating
resolution so that the quality to time ratio is maintained.
We test the developed framework in different real-time applications.
First, we present object tracking using block-based motion estimation.
We prove through this research that more reliable object tracking results
can be obtained by combining motion characteristics with feature
tracking. Second, we present a video stabilization model using the
proposed motion estimation technique. In this application, we cornpare
the performance of the proposed block-based motion estimation
scheme to the techniques specified in the literature and hence prove
the efficiency and robustness of the framework. Finally, we tackle the
problem of low level video indexing using a weighted combination of
features during motion estimation.
History
School
- Science
Department
- Computer Science
Publisher
© Harish BhaskarPublication date
2007Notes
Doctoral Thesis. Submitted in partial fulfillment of the requirements for the award of Doctor of Philosophy of Loughborough University.EThOS Persistent ID
uk.bl.ethos.493281Language
- en