CAREER: Stochastic Models for Video-Equipped Intelligent Environments

Rensselaer Polytechnic Institute, Troy NY

Investigators

Abstract

A surprisingly large number of safety tasks can be viewed as the detection of unusual events. This research will develop video analysis systems that can distinguish ``unusual'' from ``normal'' activities without specifying the unusual events ahead of time. This stands in contrast to most current systems which are designed to recognize special, pre-determined events. Three major thrusts of work will make the development of such video safety systems more realistic: (1)The PI will significantly increase the precision of parametric process models by including the event duration and the time of an event. (2)These complex models require more training data, yet only few data points will be available to estimate the time-dependent parameters. Therefore, this research will explore several different ways of estimating accurate models robustly from limited data. (3)Finally, this research will investigate how to use HMMs to detect deviations of the current time-series from the learned model. The theoretical work will be put to practice on a task that will increase the safety of independently living senior citizens. The system will learn a model of the daily habits of the person based on video data from multiple cameras in the apartment. Emergency situations like strokes and falls can then be detected as deviations from the regular behavior, and a care taker can be notified. In the later stages of the project, the new models will also be applied to other safety and traffic tasks. The education component of this award includes the development of extensive tutorials on stochastic modeling techniques which will be made available to the public online, as well as an interdisciplinary seminar on stochastic modeling at RPI.