Classification
of vehicles in points clouds of urban scenes
($45,500)
March 2015
Description
The photorealistic modeling of large-scale
scenes, such as urban structures, has received signicant
attention in recent years. This is a challenging problem as
urban environments are a mixture of buildings, people, vehicles,
street level structures, roadways, curbs etc. The complexity of
urban environments has to do with the variability of objects,
partial visibility and occlusions, and varying object
resolution. A major goal is the photorealistic rendering of such
scenes for inclusion in products such as Google Map or Google
Earth. A signicant obstacle in that direction is that many
objects are only partially sensed. Vehicles represent a major
class of objects in this category. Locating the vehicles in the
point cloud, identifying their pose and type is an
important step towards their complete photorealistic
representation.
Financial Mathematics is a branch of
applied mathematics based on stochastic analysis and optimization
that has gone through a period of extensive growth over the last
years. Originally concentrated in portfolio management and
derivatives pricing, the use of sophisticated mathematical methods
has grown to a wide array of different applications in finance. This
workshop will focus on three topics of current interest in the area
of financial mathematics: high frequency trading, optimal investment
under transaction costs, and systemic risk. All of the above areas
have received great attention in recent years, and a significant
number of open problems emerged in each of these topics. The
objective of our workshop is to familiarize students with these
topics and present to them some of the open problems as well as
hands-on guidance on possible solutions. The professional
development component of the workshop will shed new lights on the
various possibilities of a career in the area of financial
mathematics, both in academia and in industry.
People
Co-PIs
Maxim Bichuch, Michael
Carlisle, Birgit Rudloff, Stephan Sturm
NSF - DMS - ATD (2012-2015)
(Brooklyn College: $278,154)
Sequential
detection and identification of multiple co-dependent
epidemic outbreaks
Description
This project is key to the development of next generation
quantitative algorithms for detection of epidemic outbreaks.
The investigators address two focus problems that arise in
epidemic surveillance, namely that of quickest detection of (a)
spatially and (b) pathogen heterogeneous outbreaks. An early and
accurate response is achieved by taking advantage of the
co-dependent nature of the corresponding syndromic observations
and by appropriate modeling of this dependency. To this end,
the investigators develop innovative online quickest detection and
sequential classification techniques to analyze multiple
correlated data streams undergoing distinct changes. These
techniques are assessed through their ability to optimally issue
timely outbreak alerts with minimal false alarm rates. Moreover,
the investigators address the problem of early detection and
identification of an epidemic outbreak by designing a simultaneous
min-max change-point detection and classification algorithm of a
single data stream with unknown post-disorder
characteristics. In this way, the investigators are able to
also address the problem of model uncertainty and build robust
algorithms. Finally, the investigators combine their expertise by
carrying out a multi-faceted comparison of alternative
formulations (especially Bayesian versus min-max) for the focus
problems, thus creating a model-free state-of-the-art toolkit
targeting highly complex bio-surveillance data.
Sequential
Classification and Detection via Markov Models in Point
Clouds of Urban Scenes
Description
One of the most important problems in 3D computer vision and
graphics is the automatic scene reconstruction from 2D and 3D
images. Recently, the reconstruction of complex urban scenes has
attracted significant interest. This is because accurate 3D city
models are paramount in the further development of a variety of
fields such as urban planning, architecture, and archeology. They
are also very important for applications commonly used in everyday
life such as street map visualization and navigation, as well as
in the film and construction industries. Automatic 3D image
reconstruction and classification of urban scenes, though, is a
problem whose complexity still challenges today's research
community. 3D reconstruction of city models is achieved through
data acquisition using a variety of devices such as laser scanners
and regular cameras. While laser scanners provide dense, detailed
and accurate 3D points, they suffer from slow speed which
dramatically increases the cost of acquisition. For more
information please go to link.
Sequential Detection and
Classification in 3D Computer Vision
Description
The problem of quickest detection and classification in the
statistical behavior of sequential observations is a classical
one, with numerous applications in engineering, economics and
epidemiology. In today's fast-growing technologies new areas of
applications constantly emerge. In particular, the automatic 3D
image reconstruction and classification of urban scenes is a
problem whose complexity still challenges computer scientists. It
has traditionally been treated through the acquisition of data
using laser-scanners, which produce high-resolution images, but
can be very slow. It is thus essential to concentrate laser
scanning only to the areas of interest, which leads to fast
decision-making about areas of interest. This can save significant
time and cost, while still producing high-resolution 3D images.
The goal of this project is to develop and implement real-time
algorithms for processing and analyzing 3D laser range data. The
high-dimensional nature of the data is reduced by a clever
innovative selection of a measurement model. Interdependent
streams of observations are then processed by on-line parametric
and non-parametric classification and detection techniques. And
finally, new statistical models are used to capture obstacles in
urban scenes. This provides a systematic treatment of the problems
of fast and efficient 3D image classification using
high-resolution laser data. For more information please goto link.
People
Undergraduate Students
Artur Sahakyan
(Brooklyn College alumnus, currently employed at IBM's
dispatching division).
NSA MSP Probability (2009-2012)
($30,000) Quickest Detection in
correlated multi-sensor systems
Young
Investigator's award
Description
This project is central to the detection and identification of
abrupt changes in sequential observations in complex multi-source
systems. The detection and identification of abrupt changes arises
in many different areas. Examples of these areas are the detection
of enemy activity, quality control, the detection of intrusions in
computer networks and signal detection from multiple sources such as
wireless communications. Although the classical problem of quickest
detection has been treated in many forms in the literature dating
back to the 1930s, the challenges presented by today's fast-growing
technologies cannot be properly addressed by the traditional
techniques. We intend to address these challenges using a
combination of mathematical tools drawn from probability, modeling,
stochastic processes and partial differential equations. This is a
comprehensive project whose solution will involve the synergy of a
variety of mathematical tools. The research proposed will not only
present novel methods of incorporating dependencies across channels,
but could also potentially transform the systems used in defense,
target detection, wireless communications, portfolio management and
intrusion prevention of attacks in networks.