Robust tracking of appearance by sparse template adaptation

The problem of real-time tracking appearances in dense visual clutter is challenging. This paper provides a solution to the tough problem by combining an adaptive appearance model and a template-based condensation tracker. As an adaptive appearance model, we utilize the WSL appearance model proposed by Jepson et al. We adopt the sparse template condensation proposed by Shakunaga et al. as a framework of real-time tracking. Rapid and reliable tracking is performed in the framework by matching only a few percent of the whole pixels in the template. Since a sparse template is made up from the stable component of the WSL template, which is called the S-template in this paper, the WSL model and the sparse condensation tracker can collaborate with each other through the sparse S-template. The tracker works as almost rapidly as the sparse template condensation since the WSL model efficiently facilitates the template adaptation.