Modern Statistical Techniques For Computer Vision

Rutgers University New Brunswick, New Brunswick NJ

Investigators

Abstract

The power of modern statistical techniques is exploited to develop novel approaches to fundamental problems in image understanding. The common statistical foundation for a broad class of vision tasks is identified first, and then the underlying key problems are solved in a rigorous framework. For example, since heteroscedasticity inherently appears in most 3D vision tasks the development of a complete estimation procedure (which includes imposing further geometric constraints on the parameter estimates) is of great importance for computer vision. Such a procedure can provide a faster and more reliable alternative to the widely used (and too general) nonlinear Levenberg-Marquardt method. The task of deriving the 3D description of a scene (static or dynamic) from an image sequence captured with an uncalibrated camera was chosen as a testbed. Estimation problems related to self-calibration, object recognition supported by uncertainty information, will not only allow us to enhance our existing toolbox but also to gain expertise in integrating these tools in a closed-loop autonomous vision system.