Activity analysis in
intelligent rooms
The
system is able to track people in 3D and is aware of their identities,
and of the identity of the current speaker. Based on this information
and the known environment, an important set of activities can be defined
and recognized in real time. 3D tracking is based on the video from static
cameras with highly overlapping fields of view. Active camera network
is used for two purposes: taking snapshots of people's faces to be used
for face recognition and for capturing video of interesting events. People
are recognized by combining face and voice recognition results to achieve
a more robust performance. In addition to the described real-time functionality,
the system enables reviewing of past events in the environment. The events
are summarized graphically to enable the user to easily grasp the spatio-temporal
relationships between events and people that participated in them. This
graphic also serves as a user interface for an interactive review of the
events such as replay of video associated with a certain event.
|
AVIARY video: Digivation Conference,
September 2000 |
System demonstration CVPR Workshop
on Human Motion, Analysis and Synthesis, June 2000 |
|
Multimodal and audio analysis
Using both audio and video
sensor data, a robust performance can be achieved. We have pursued research
in audio and multimodal signal analysis.
Televiewing
Using multiple omnidirectional
video streams, we have developed algorithms for generating virtual views
from arbitrary viewpoints as well as virtual walkthroughs. Also, we have
developed techniques for seamless merging of high-resolution rectilinear
images with low-resolution omnidirectional images.
Semantic event databases
store abstracted past states (events) of the intelligent environment and
active sensory networks. These can be flexibly queried using a powerful
language for characterizing complex activities as the spatio-temporal
compositions of semantic activities. Please view this demo page.