From: Christian Shelton
To: all-ai@ai.mit.edu
Subject: GSB Talk today 5:30pm in 7th floor playroom
Date: Fri, 16 Apr 1999 13:56:43 -0400 (EDT)

******TODAY*******, Friday April 16, 1999

In the continuing Barley and Malt Seminar Series (AKA GSB),

Speaker: Half Pint From: The Girl Scout University

Title: Beverage Consumption Tracking in Multiple Camera Views using Gaussian Processes

Time: 5:30pm in the 7th floor "lounge"

Abstract:

The problem of monitoring beverage usage and consumption has gained interest among researchers due to the increased availability of statistical data and processing methods and the wide-spread use of multi-camera vision systems. These have allowed mathematical techniques to be applied to a previously ill-posed problem, making exact or approximate answers tractable.

The usefulness of a system to track drink distribution is obvious. The "holy grail" of this field for many years has been a system which can automatically answer queries such as "Who drank the last beer?" or "How much beer will be needed for the party tonight?"

Although these "super queries" are still outside of the reach of today's systems, we will demonstrate a promising novel approach which can answer simpler questions such as "Is there any beer in the fridge?" or "When was the fridge last stocked?" To answer such queries, we track the movement of beverages into and out of a fridge from multiple vantage points. By analyzing the statistics of the motions using Gaussian Processes, we are able to handle occlusions (e.g. removing two beers from the fridge, one behind the other) robustly. Furthermore, from multiple cameras, we can reconstruct the full range of beverage motions thus enabling beverage identification for accounting purposes and outlier-detection for use in a drink depletion warning system.

In the first part of the talk we will describe the underlying mathematical models and demonstrate their perturbation sensitivity and the running bounds on the algorithms. The second part of the talk will consist of a demonstration of the system which can keep track of the number and content of up to fifty types of beers in real time while answering questions of the form "how much of beer XX is left?" and "how long has beer XX been in the fridge?"

Note: Professor Pint will be around all day. If you have interest in his research, please e-mail drinks@ai to be placed on his schedule. General research questions can be put to him after his talk at this week's

G I R L S C O U T B E N E F I T 7ai, 5:30pm