Layered Active Contours for Tracking

This paper deals with the task of object tracking in the presence of occlu- sions and clutter by fitting a layered appearance model to data. Four major problems must be overcome: (1) the association of each pixel to a particu- lar layer (layer segmentation), (2) the determination of layer support, (3) the determination of layer appearance, and (4) determination of layer location. Tao, Sawhney, and Kumar successfully proposed a generalized expectation maximization algorithm solving these problems by directly inferring masks representing layer segmentation in conjunction with a deforming elliptical shape prior defining layer support. We extend their work with the introduc- tion of active contours: instead of directly inferring these masks, we evolve a series of curves to obtain a layer segmentation. These curves provide a natural shape prior by constraining segmentations to a family of curves local to layer supports and allow for non-rigid layer deformations through the pre- diction of unobserved appearance information during inference. A benefit of this extension is the ability to track through massive occlusions and clutter, as demonstrated on a series of difficult real-world video sequences.