Combining online feature selection with adaptive shape estimation
Adam Gatt, Sebastien Wong, David A. Kearney · 2010
The online selection of features is widely cited to provide benefits in video tracking applications where the foreground (object) and background (clutter) models must to adapt to changes in the scene. It is commonplace amongst researchers in the field to assume a rectangular bounding box between foreground and background that is known a priori and is invariant during the tracking process. In many important tracking scenarios these assumptions do not hold. Recent work has shown that online feature selection is critically dependant on the selection of the boundary between the foreground and background. In this paper we present a tracking system that uses highly accurate shape estimates generated by the Shape Estimating Filter algorithm, which probabilistically estimates object shape using both top-down and bottom-up Bayesian methods. We show that the use of these shape estimations for object modelling improves feature selection over a simple bounding box scheme.