Page 23 - EURECOM - RA2011GB

Basic HTML Version

multimodal processing techniques to detect enriched concepts in video
ideoSense aims to use innovative
multimodal processing techniques
to detect and recognize enriched
concepts (including static, dynamic and
emotional aspects) in video sequences.
Targeted applications include advertisement
selection or video recommendation based on
the content of videos being watched by users.
The project builds on the expertise of partners
specialized in audio, image, video and text
processing to enhance existing descriptors. It
uses pivot languages to process multilingual
closed-captions, active learning techniques
to limit dependency on annotated learning
corpus, efficient detectors for static, dynamic
and emotional concepts, as well as fusion
mechanisms to reinforce classifiers’ results.
In this project, EURECOM is bringing its
expertise in video sequence analysis and
indexing by developing aspects related to
Beyond gaming!
he RGBD (Red, Green, Blue, and Depth)
Kinect camera, originally conceived
to allow Natural User Interaction with
game consoles and PCs became the biggest
commercial success of the year 2011, with over
10 million units sold in less than five months.
Thanks to the individual efforts of Prime-
sense, Microsoft, and a team of hackers who
founded the project OpenKinect, a series of
hobbyist and researchers have started using
the new sensor to increase the potential of
many different kinds of applications.
Here at EURECOM, we are exploring the use
of Kinect for Biometric and Surveillance appli-
cations in the context of projects like VideoID
and ActiBio.
The possibility of sensing the 3D environ-
ment that surrounds the camera and the ability
of the Kinect to track someone’s body parts
empowers an automatic system that we deve-
loped to extract anthropometric measures.
Those measures are then used to extract Soft
Biometric information from a distance.
Based on this information we can estimate
the weight and gender of a subject in front of
the camera. To do so, we use a statistical model
built over the information extracted from the
NHANES database, a large American medical
database containing the records of more than
27,000 people.
The different approaches, whether in quan-
tity or quality offered by the Kinect capabilities
span many domains of application. With the
collaboration of the Centre for Space Human
Robotics of the Italian Institute of Technology
in Torino, we tried to solve a practical problem
faced by cosmonauts once in space. The lack of
gravity activates body mechanisms that cause
severe bone and lean body mass losses. This
forces cosmonauts to follow a regimen of exer-
cises and diet and to monitor their body mass,
which is a problem considering the weightless
conditions. This is why a system that could
visually estimates their mass is interesting.
The preliminary experiments we conducted to
explore the possibility of applying our research
outcomes to this problemhave shown that it is
possible. The performance is already close to
the systems currently on board the Internatio-
nal Space Station.
Other possible applications include, but
are not limited to the possibility of recognizing
someone based on his/her physical measure-
ments, or using the Kinect as a medical sup-
port for telemedicine thanks to the capabilities
enabled by our automatic weight estimation.
content representation through spatial-
temporal descriptors, and usual common
classifier adaptation to these new descriptors.
Project progress is assessed with a video
created by our industrial partner from an
existingWeb service. This enables us to
evaluate the impact of our algorithms on
a real application. The project also uses
the TRECVID campaign as a state-of-the-art
benchmark. The algorithms developed in the
project are readapted for enhanced efficiency
in computing resources, and implemented in
the production environment of the industrial
partner’s video server. The partner will
therefore benefit from these technologies in a
real Web service and improve its international
competitive edge.
Thesis advisor
Jean-Luc Dugelay
University of origin
Politecnico di Torino
multimedia communications
young researcher
Recommended videos
The video you arewatching
Suggested ads
2011annual report