Robust object tracking via active feature selection

Description

In this project, we proposed a robust tracker based on an online discriminative appearance model. In order to design a robust appearance model, we developed an online active feature selection (AFS) approach via minimizing a Fishier information criterion. We showed that the features selected by our proposed online AFS boosting algorithm are much more informative and discriminative than those selected by online MIL boosting algorithm which maximizes a likelihood loss function. The AFS appearance model can well handle large appearance changes. Numerous experimental results and evaluations on challenging video sequences demonstrated that our AFS tracker outperforms other state-of-the-art algorithms in terms of efficiency, accuracy and robustness.

Some sampled tracking results on Pedestrian sequence

reference : 

Zhang, Kaihua, et al. “Robust object tracking via active feature selection.” IEEE transactions on circuits and systems for video technology 23.11 (2013): 1957-1967.

Regularized Robust Coding for Face Recogntion

Robust Kernel Representation with Statistical Local Features for Face Recognition