Run-time Dynamic Model Estimation for a Multi-hypothesis Mean Shift Tracker
Tom Caljon, Peter Schelkens · 2005
We resort to a mean-shifting color-based particle filter for the pur-pose of tracking objects in a video sequence. Primary use case is the creation of interactive versions of video sequences i.e. video sequences containing clickable objects. As the tool needs to be used by non-technical users specification of parameters should be kept to a minimum. For objects types that were not anticipated and for which a trained dynamical model is not available, we then pro-pose run-time estimation of the dynamical model instead of sim-pler models such as constant velocity with fixed noise variance. We present results on the performance of such tracker and com-pare with non-meanshifting particle filters and a single-hypothesis mean shift tracker. 1.