Statistical and geometric modeling of spatio-temporal patterns for video understanding

摘要: Title of Dissertation: Statistical and Geometric Modeling Spatio-Temporal Patterns for Video Understanding Pavan Turaga, Ph.D. Oral Examination, 2009 Dissertation directed by: Professor Rama Chellappa Department Electrical Computer Engineering Spatio-temporal patterns abound in the real world, understanding them computationally holds promise enabling a large class applications such as video surveillance, biometrics, computer graphics animation. In this dissertation, we study models algorithms to describe complex spatio-temporal videos wide range applications. The pattern recognition problem involves recognizing an input instance known class. For problem, show that first order GaussMarkov process is appropriate model space primitives. We then primitives not Euclidean but Riemannian manifold. use geometric properties manifold define distances statistics. This paves way temporal variations these techniques activity discovery from long videos. on other hand, requires uncovering datasets unsupervised manner automatic indexing tagging. Most state-of-the-art index according global content scene color, texture brightness. discuss based examine various issues involved effort general framework address problem. design cascade dynamical systems clustering their dynamics. augment traditional two ways. Firstly, activities systems. significantly enhances expressive power while retaining many computational advantages using models. Secondly, also derive methods incorporate view rate-invariance into so similar actions are clustered together irrespective viewpoint or rate execution activity. learn parameters stream demonstrate how given sequence may be segmented different clusters where each cluster represents Finally, broader impact tools developed dissertation several image-based problems involve statistical inference over non-Euclidean spaces. geometry underlying leads more accurate than approaches. present examples shape analysis, object recognition, video-based face age-estimation facial features ideas.