"Estimation Theory Meets Visual Tracking and Active Contours"

Patricio Vela, School of Electrical and Computer Engineering, Georgia Tech

Traditional closed-curve, segmentation-based tracking algorithms perform tracking by solving temporally decoupled optimizations over a given image sequence. When the underlying models of the segmentation objective function fail to hold due to their simplicity or due to imaging noise, the segmentations fail to be consistent. Introducing temporal filtering is not straightforward since the space of (smooth) closed curves forms an infinite-dimensional space. Approaches relying on finite-dimensional approximations have been attempted, however they hold when the object shape variation can be restricted to a low-dimensional, parametrized contour representation. Particle filter approaches exhibit poor computation efficiency due to the need to sample from a large space.

This talk presents two simple approaches that seek to respect the manifold structure of shape by filtering on implicit representations for shape. The first approach introduces a local coordinate system for curves. Within this coordinate system, filtering behaves similarly to the finite dimensional case under certain smoothness conditions that typically hold. The second approach explicitly injects uncertainty into segmentation process and examines how the segmentation decision boundary is impacted by the uncertainty. Filtering on the space of density functions follows from an attempt to minimize this impact. The two approaches currently rely on a first-order, zero velocity assumption for the dynamics, however extensions to the second-order, zero-acceleration case are currently being considered.

Patricio Vela is an Associate Professor in the School of Electrical and Computer Engineering at the Georgia Institute of Technology. His interests lie at the interface of control and dynamical systems theory and computer vision. In particular, he studies the role of the feedback loop within the context of visual tracking, whether it be for control or estimation purposes. This particular research interest stemmed from his graduate work in geometric control for biologically inspired locomotive systems and his post-doctoral work in controlled active vision.

Unless otherwise stated, the content of this page is licensed under Creative Commons Attribution-ShareAlike 3.0 License