Cascaded generative and discriminative learning for visual tracking

Abstract

We propose a novel visual tracking framework which incorporates a generative and a discriminative tracker in a cascaded manner for robust visual tracking. The generative tracker filters out most easy candidates in the early stage and retains a few most confusing samples. The discriminative tracker then re-evaluates these samples using the Partial Least Squares (PLS) discriminant analysis. Both trackers are collaboratively updated online to adapt to appearance changes during tracking. The proposed approach explicitly learn the appearance difference between the target and the most confusing distracters and is thus able to alleviate the "drifting" problem. Comparing tracking performances on challenging video sequences, which contain significant appearance changes, severe occlusions, out of the field-of-views and cluttered backgrounds, demonstrate the promising of the proposed method with respect to recent state-of-the-art trackers. © 2013 Springer-Verlag.