Detecting People in Crowds by Bayesian Clustering
Speaker: Gabriel Brostow , University of Cambridge
Date: May 25 2005
Time: 2:45PM to 3:45PM
Location: G449 (Patil/Kiva)
Host: Greg Shakhnarovich, CSAIL
Contact: Greg Shakhnarovich, xx38170, gregory@csail
Relevant URL: ABSTRACT
The motion of a crowd can be characterized by the velocities
of individuals and the overall density of bodies. While a good
higher level crowd-tracking system should take advantage of human
appearance and motion models, this talk describes a decidedly
simple low level Bayesian clustering algorithm with the same goal.
We track image features and probabilistically group them into
clusters representing independently moving bodies. The numbers of
clusters and the grouping of constituent features are determined
without supervised learning or any human-specific model. The new
approach is instead, that space-time proximity and trajectory
coherence through image space are used as the only probabilistic
criteria for clustering. An important contribution of this work is
how these criteria are used to perform a one-shot data association
without iterating through combinatorial hypotheses of cluster
assignments.
In the challenging domain of complex crowd footage filmed with a
stationary uncalibrated camera, this talk examines the
implementation and experiments of our simple framework. Our
technique could serve as preprocessing and training
for other systems, or can stand on its own as a people counter.
Joint work with Roberto Cipolla
See other events that are part of MIT Machine Vision Colloquium 2004/2005
See other events happening in May 2005
