Robust visual tracking with classifier-like appearance model and entropy particle filter

Abstract

The detection based visual tracker treats tracking as the object and its surround background online classification problem. There are two main difficult issues in this method: one is to specify exact labels for the online samples, the other is to avoid template drift that caused by wrong update of the classifier-like appearance model. To overcome the problems, a novel tracking algorithm based on online Multiple Instance Learning (MIL) and entropy particle filter is proposed. Main contributions of our work are: (1) we introduce MIL in particle filter visual tracking framework to reduce the online training error of the classifier-like appearance model; (2) the appearance model consists of an initial fixed MIL classifier and an online dynamic MIL classifier; (3) a particle set maximum negative entropy criterion is designed to online fuse the two classifiers. Experimental results verify the effectiveness of the proposed algorithm.