Non-Parametric Representations of Motions and Actions in Video

Washington University, Saint Louis MO

Investigators

Abstract

The PI proposes to establish a 3 year program to develop non-parametric representations of the behaviors and actions of unknown objects in video sequences. The approach is to use the generative statistical modeling tools of Pattern Theory to optimize non-linear dimensionality reduction tools to be applied to natural image sequences. Applications include registration and de-noising of medical or aerial image data sets, understanding of non-rigid and deformable motion, and content based video indexing. The intellectual merit includes formal tools to indicate under what real world conditions techniques such as Isomap fail. Collaborators whose focus is the study of cardiovascular MRI data, diagnostic human motion analysis, and the mechanics of head injuries facilitate explicit testing of these approaches in real and diverse applications. The broader impacts of this proposal include expanding access to healthcare by allowing new experimental protocols for MRI imaging studies, and gives tools that support human motion analysis in unstructured settings. Extracting general representations from video supports content based indexing which lies at the heart of the multi-media component of the national digital library initiative. A web-based portal (Wumap) will serve as a national resource for automatically generating web-based visualizations of natural patterns and will have portions designed for young K-12 students.