September 12, 2012
Title: Physics in Human Motion Estimation and Scene Understanding
Abstract: Recent work has demonstrated that a physically realistic model of the world can greatly aid in estimating human motion from video. While improved tracking and estimation methods are one natural application of physics based models of human motion there are other interesting applications. This talk will present two applications of physics-based models of human motion to the broader context of dynamic scene understanding. From motion alone we are able to recover scene properties, such as the position of a ground plane and the orientation of gravity. We are able to do this using both noisy tracking results and motion capture data. Further, when given noisy motions such as those typically generated by modern tracking algorithms, we are able to restore a higher quality, physically realistic motion. When data is missing (such as entire frames or body parts which were not tracked) we are able to recover this missing data in a physically plausible manner.
Biography: Marcus Brubaker received his BSc, MSc and PhD in Computer Science from the University of Toronto. His graduate work was performed under the supervision of David Fleet. Since 2011 he has been a postdoctoral scholar at Toyota Technological Institute at Chicago and has worked jointly at the University of Toronto and, currently, the Toronto Rehabilitation Institute. He also consults with Cadre Research Labs on machine learning and computer vision related projects.
He has worked on video-based human motion estimation, physical models of human motion, video-based tracking, Markov Chain Monte Carlo methods, ballistic forensics, electron cryo-microscopy, time series modelling and automatic vehicle localization. His interests span computer vision, computer graphics, machine learning and statistics.