A Robust Online Learning Method for Object Tracking

Abstract

In this paper, we present an robust object tracking method to deal with constantly appearance changing and partial occlusion. Our method is based on particle filtering framework which allow us to handle partial occlusion over a few frames as well as few false predictions caused by rapid appearance changing. An online multiple instance learning algorithm is embedded in the particle filtering framework to adaptively learn the changing appearance model of the object and background. Weighted sampling method and synthesized extension are introduced in the learning process to make the training model robust and avoid over-fitting. Experimental results prove that the proposed method outperforms the other state-of-the-art tracking via online boosting algorithms.