Within the wide
area of databases, my students and
I are mainly interested in spatio-temporal data
management, query processing within sensor networks,
and
content-based image retrieval. Some of the
publications related to my research can be found via this
page on DBLP (thanks for Michael Ley) or
via this
page (both pages contain publications related to my previous
research as well).
Spatio-temporal Data Management
Spatio-temporal
data has always existed, but is becoming more common nowadays,
therefore generating demand for more effective and efficient
data management techniques. At a very
high level, the task at hand is managing with who was where
and when.
A few sample application scenarios are as follows.
A rental car companies can track their fleet using GPS devices
installed in the vehicles in order to verify whether the rental
contract ws violated.
Cell phones can also be equipped with GPS
devices which, for instance, would facilitate the location of the
person carrying it when an emergency call is placed. Animals behaviour
(or changes thereof) can also be seen as spatio-temporal data. In
particular, one could correlate changes
of behaviour to changes in the environment.
Another application could be to proactively advise drivers of
current road accidents based on their usual driving routine.
The following list isa sample of the issues we are working (jointly
with Prof.
Sander) within this domain.
- Most indices proposed in the literature cannot be embedded
within existing RDBMS, therefore not solving the indexing problem from
a practical perspective. We are investigating how to map the indexing
problem in order to use an off-the-shelf RDBMS to efficiently query
historical spatio-temporal data.
- Most indices for spatio-temporal data are based on the classic
R-tree, and it is well known that the node splittling policy is of
great
important with respect to query performance. We are working on a
provably
optimal node split policy for trajectory data as well as near-optimal
but highly efficient split policies that can be used virtually within
any MBR-based index.
- When observing the movements of objects in many spatio-temporal
domains
one can observe that the movements are hardly random, i.e., patterns
exist. We are researching the ideas of identifying such patterns and
subsequently exploring the obtained ones in order to proactively
suggest alternative courses of action.
Data Management in Sensor Networks
A close and relatively new research topic related to spatio-temporal
data is that of data management on (ad-hoc) sensor networks. For
instance, using this paradigm, (very small) sensors can be spread over
a large area (e.g., a forest) in order to gather and store data which
can be used for (a posteriori) query processing. A chief concern in
this environment is to minimize the energy consumption during the
network's lifetime, in particular during query processing time.
Some of the research topics we are currently working on are the
following.
- Given the very nature of the sensors (e.g., small size, fragile
structure and limited battery lifetime) it may be unfeasible to assume
the network topology to be known at all times. We are working
on efficient query routing techniques that optimize the networks
lifetime. The fact that some queries may be interested only in
aggregated data offer opportunities for further optimization of such
techniques.
- A non-trivial variation of the problem above is when the sensors
are
also able to move. There are now two spatial aspects to the problem,
of the measured data and of the sensors, the coordination of these,
as well of the temporal dimension require novel techniques, which we
are
currently working on.
Content-based Image Retrieval
Another area I
have been working on is that of content-based image
retrieval. My main concern in this domain is how to abstract image
data in order to efficiently store and effectively query them.
Typically a few simple, but fairly standard features, such as color and
texture, along with vector-based similarity queries suffice to provide
good results. Most of my work has assumed that a well-design linear
scan of compact image information can be at least as effective and more
efficient than using complex features and tree-based indices. Some of
the research results we have obtained are summarized next.
- It is known that humans do not react linearly to the increase of
some
stimuli. We have taken advantage of this by using non-linearly
discretized color histograms to abstract images. This resulted in
greatly reduced metadata which could be fit in main-memory (e.g.,
thousands of images using a few megabytes) and searched without the
need of any indexing structure.
- Further exploring the idea of discretizing color histograms we
have
devised a novel histogram-like image abstraction that, although only
concerned with the color of the pixels, could also capture texture
information, improving even further query effectiveness.
- Many times users are interested in finding images that contain a
given
query image. We have explored this problem, called sub-image image
retrieval, and have experimented the use of relevance feedback
techniques to improve the query effectiveness with the help of the
user.
- Some similarity measures do not lead to vector data but use a
well-defined metric distance. We have worked on improving existing
techniques for metric indices (e.g., by taking advantage of how
efficiently one can execute a linear scan on secondary storage) as well
as designing a new access structures which explores the notion of
clustering and dimensionality reduction in order to speed up query
processing.
- Work on indexing non-metric is needed as well. Not only some
effective
similarity measures are based on non-metric distances but it is also
argued that the metric axioms are non-practical (despite their
theoretical usefulness). Given the large number of approximation used
in this domain, exploring approximate answers seems a feasible venue
for research.
Research Support
My research has been mainly supported by NSERC (through individual
and equipment
research grants) and Canadian
Heritage
(through the New Media Research Networks Fund program).
