A Closed-Loop Filtering Framework for Active Contours

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

The project described in this proposal seeks to a develop a framework for generating robust computer vision algorithms for use in closed-loop systems. A principal goal of the research endeavour is to examine the role that control theory may play in enhancing closed-loop computer vision algorithms. Intellectual Merit. As a sensor, the imaging system can be wrought with noise, either through the actual sensing process or through the geometry of the imaged scene. Consequently, the computer vision process can be interpreted as a signal processing task in the presence of noise and uncertainy. This is further compounded for closed-loop vision systems, because the control effort can induce additional disturbances. Through an analysis of the classical Luenberger observer for finite- dimensional systems, the PI proposes to systematically build up a similar framework for filtering of closed curves, which are the by-products of th e computer vision algorithms known as active contours. Essentially, this research endeavour involves the development of a Kalman filter algorithm for closed curves, and involves a principled predict and update scheme. The main investigative challenge lies in the fact that closed curves form an infinite-dimensional space which classical state-space observer theory is not capable of handling. Broader Impact. The successful consideration of computer vision algorithms as components of a control and dynamical system has the ability to tremendously increase their level of robustness, without severely complicating the nature of their processing. As a particular application demonstrating the potential impact and challenges of this research avenue within a specific area, the PI seeks to study the closed-loop control and estimation of bio-membranes at the mico-scale. The observer concept described herein has the capacity to improve the signal generated from vision-based algorithms that form the main information pathway of a closed-loop process. The proposed education plan incorporates many of these ideas into existing computer vision courses, as they are essential components in the use of computer vision for feedback-driven systems. Secondly, summer research and senior-design projects are anticipated to motivate the importance of control and dynamical systems theory for computer vision.