Monday, March 8, 1999
3:00 PM (refreshments 2:45)
Room NE43-941
EECS Special Seminar
Abstract
Existing computer interfaces provide rich audio-visual output but are impoverished in terms of input--they are literally deaf and blind to their users. Perceptual interfaces offer the promise of computers understanding their users through direct observation of user expression, posture, and motion. This talk presents new techniques for the visual tracking of people and understanding their actions in cluttered or noisy environments. First, I'll describe a new method for finding point correspondences in images with hands, faces, and other articulated bodies. This method succeeds in many important and common real world cases where previous methods regularly failed. The key is a new image transform which captures the shape of local image homogeneity and makes correspondence unambiguous. Second, I'll show an application which combines stereo depth estimates with results from color and appearance modules to construct a real-time interactive video display. It senses the position and number of users observing it, adjusts displayed images accordingly, and robustly recognizes users returning from out of view, based on characteristics from all three visual modalities. Finally, I'll show how to track object pose within this framework, exploiting linear rigid-body motion constraints on registered intensity and video-rate stereo depth observations.
Hosts: Professor R. Brooks and Professor E. Grimson
|
Modified: May 3, 1999|
Current events|
Your comments
and inquiries are welcome.