Nonlinear methods for parametric grouping and modeling of motion

This project investigates unsupervised learning of motion models from visual sequences.

Recovery of detailed motion models would enable numerous applications of motion recognition, analysis and synthesis in areas such as surveillance, monitoring, multimedia, as well as understanding of motion in general. Describing motion is difficult because it requires the knowledge of motion categories; however, their manual specification is laborious and automated extraction is aggravated by noise and ambiguous measurements of the human motion from video. The central goal of this project is to develop methods and algorithms for grouping and analysis of motion acquired from video sequence using a unified statistical framework. The framework relies on concepts from statistical modeling, Bayesian networks (switching linear dynamic systems) and probabilistic reasoning to develop motion models of sufficient detail, quality, scalability and robustness that could be used to (1) group motion into perceptually plausible categories and styles, (2) recognize and predict motion dynamics, and (3) possibly synthesize natural-looking motion. The utility of the framework will be evaluated on sequences of human body or hand motion. The novel methods and algorithms developed in the course of this project may have wide applicability in many areas that involve modeling and grouping of sequence data, such as biotechnology and life sciences.