Visual attention models for far-field scene analysis

The amount of information available to an intelligent monitoring system is simply too vast to process in its entirety. One way to address this issue is by developing attentive mechanisms that recognize parts of the input as more interesting than others. We apply this concept to the domain of far-field activity analysis by addressing the problem of determining where to look in a scene in order to capture interesting activity in progress. We pose the problem of attention as an unsupervised learning problem, in which the task is to learn from long-term observation a model of the usual pattern of activity. Such a statistical scene model then makes it possible to detect and attend to examples of unusual activity. We present two data-driven scene modeling approaches. In the first, we model the pattern of individual observations (instances) of moving objects at each scene location as a mixture of Gaussians. In the second approach, we model the pattern of sequences of observations -- tracks -- by grouping them into clusters.We employ a similarity measure that combines comparisons of multiple attributes -- such as size, position, and velocity -- in a principled manner so that only tracks that are spatially similar and have similar attributes at spatially corresponding points are grouped together. We group the tracks using spectral clustering and represent the scene model as a mixture of Gaussians in the spectral embedding space. New examples of activity can be efficiently classified by projection into the embedding space. We demonstrate clustering and unusual activity detection results on a week of activity in the scene (about 40,000 moving object tracks) and show that human perceptual judgments of unusual activity are well-correlated with the statistical model. The human validation suggests that the track-based anomaly detection framework would perform well as a classifier for unusual events. To our knowledge, our work is the first to evaluate a statistical scene modeling and anomaly detection framework against human judgments.

Department dc:contributor.department

Massachusetts Institute of Technology. Dept. of Electrical Engineering and Computer Science.

Grantor dc:publisher

Massachusetts Institute of Technology

Year dc:date.issued

2007

Author dc:creator

• Ižo, Tomáš, 1979-

Advisor dc:contributor.advisor

• W. Eric L. Grimson.

• Electrical Engineering and Computer Science.

Statement dc:rights

• M.I.T. theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission. See provided URL for inquiries about permission.

Licence dc:rights.uri

• http://dspace.mit.edu/handle/1721.1/7582

Language dc:language.iso

eng

Handle dc:identifier.uri

http://hdl.handle.net/1721.1/40314

OAI identifier oai:identifier

oai:dspace.mit.edu:1721.1/40314