| Abstract | The ability to automatically detect activities invideo is of increasing importance in applications
such as bank security, airport tarmac security, bag-
gage area security and building site surveillance.
We present a stochastic activity model composed
of atomic actions which are directly observable
through image understanding primitives. We focus
on answering two types of questions: (i) what are
the minimal sub-videos in which a given action is
identified with probability above a certain thresh-
old and (ii) for a given video, can we decide which
activity from a given set most likely occurred? We
provide the MPS algorithm for the first problem,
as well as two different algorithms (naiveMPA and
MPA) to solve the second. Our experimental re-
sults on a dataset consisting of staged bank robbery
videos (described in [Vu et al., 2003]) show that
our algorithms are both fast and provide high qual-
ity results when compared to human reviewers.
|
