ABSTRACT
LK matching typically employs simple brightness constancy assumptions and uses Sum of Squared Difference (SSD). We chose to base our tracking on MI because of its robustness to environmental lighting/noise, pronounced maxima and similar computation cost to SSD. Our earliest attempt at using MI in a tracking context was the M3I tracker (Dowson and Bowden 2004) which developed into the Simultaneous Modelling and Tracking (SMAT) algorithm (Dowson and Bowden 2005) (Dowson and Bowden 2006b). SMAT was an on-line tracking algorithm that, given a single image patch in the first frame, would track and learn a hierarchical constellation model of appearance and structure on the fly. As such, it builds a model of appearance variation as it tracks, becoming more robust overtime. Tracking was performed in an optimised LK framework but using MI as the similarity measure.
