Login
Computer Science @ Rochester
Friday, March 31, 2006
11:00 AM
CSB 209
Deva Ramanan
Toyota Tech'l. Inst. at Chicago
Training a Computer to See People
One of the great, open challenges in machine vision is to build a system that can "see people". A reliable solution opens up tremendous possibilities, from human computer interaction to video data-mining to automated surveillance. This problem is difficult for several reasons. Determining the number of people present and estimating their body configurations is notoriously hard. It is also problematic to describe a person's actions because there is not a canonical set of categories into which to classify activities.

In this presentation, I will discuss our approach. We develop a tracker that learns the appearance of each person in a video. Our tracker can also be used as a general purpose system for model-building and recognition.recognition, and model-building. To recognize activities, we use an "analysis by synthesis" approach that decouples recognition from the specific choice of activity categories. We have tested our system on hundreds of thousands of frames (several orders of magnitude more than previous approaches). We show results for tracking and activity recognition on both unscripted indoor and outdoor footage, a feature-length film, and historic sports footage (from the 2002 World Series and 1998 Winter Olympics).