Multi-camera tracking via online discriminative feature and multi-cue MRF

Abstract

Visual tracking across distributed cameras with disjoint views consists of many challenges, such as illumination changing and similar appearance of multiple persons. In this paper, we present a new solution to the problem in the formulation of Multi-Cue Markov Random Field (Multi-Cue MRF), and employ the max-product linear programming (MPLP) algorithm to find the MAP configuration of MRF. Moreover, in order to bridge the gap among different camera views, we propose a hybrid strategy which integrates spatio-temporal relationship modeling, online visual feature selection and local pair-wise code (LPWC) extraction into one framework. Finally, experimental results conducted with challenging video sequences verify the effectiveness of our method. © 2013 Springer-Verlag.