Robust probabilistic estimation of uncertain appearance for model-based tracking

We present a method for appearance management for model-based visual tracking. The approach provides a concise mechanism for mapping image information to a model surface by taking into consideration uncertainties from the mapping process and by propagating appearance information from visible high-confidence surface regions to hidden or low-confidence regions. The method allows acquiring and utilizing appearance information online from available data, eliminating the need for offline appearance information. The problem is modeled with a continuous Markov random field (MRF) with clique potentials defined as energy functionals over color distributions with the Earth mover distance replacing traditional distance measures between random variables. A general MAP-MRF solution usable for general distributions as well as an efficient real-time capable linear close form approximation are presented.