Robust tracking via online Max-Margin structural learning with approximate sparse intersection kernel

Abstract

In this paper, we propose a robust visual tracking algorithm using online Max-Margin structural learning with sparse intersection kernel approximations to address the appearance variation, such as pose and scale variations, illumination changes, and occlusion. In the proposed tracking framework, we introduce a new sparse approximate HOG-based appearance model to represent the tracked object in high-dimensional feature space, which is efficient for measurement of non-linear intersection kernel. We then propose a optimisation method by extending Pegasos to the structural output with sparse intersection kernel in an online setting. Based on the new discriminative appearance models, we build a tracking algorithm using a Bayesian state inference framework to update the dynamic motion of object. The qualitative and quantitative experimental evaluations on 14 challenging video sequences demonstrate that the proposed tracking algorithm outperforms state-of-the-art trackers. We also show the low time complexity of the proposed tracker.