Abstract: The scale of video data
continues to grow exponentially, including city-scale wide-area
aerial video showing hundreds or thousands of simultaneous movers.
Extracting the most interesting, salient content from this type of
video is of increasing importance as the data volume grows while
the vast majority of events are not of interest. However,
traditional methods often fail because of low resolution and low
frame rates in this domain. At Kitware we have developed methods
for detecting events, actions, complex activities, patterns of
life and anomalies in large-scale video domains. We detect events,
anomalies and complex activities efficiently by detecting and
tracking all movers, then characterizing their behavior using
event-independent descriptors. Efficient inference is achieved
through layered, approximate evaluation as model complexity
increases. Functional scene elements such as parking spots are
recognized by analyzing behavior within and around them.
Behavioral normalcy models are learned, and anomalies are detected
using location-dependent and location-independent techniques. The
talk will provide an overview of these methods and results on
wide-area and ground-level surveillance video.