Robust face tracking with locally-adaptive correlation filtering

Abstract

A face tracking algorithm based on locally-adaptive correlation filtering is proposed. The algorithm is capable to track a face with invariance to pose, gesticulations, occlusions and clutter. The target face is chosen at the beginning of the algorithm. Afterwards, a composite filter is designed to recognize the face in posterior frames. The filter is adapted online using information of current and past scene frames. The adaptive filter is constructed by combining several optimal templates designed for distortion invariant target recognition. Results obtained with the proposed algorithm using real-life scenes, are presented and compared with those obtained with a recent state-of-the-art tracking method, in terms of detection efficiency, tracking accuracy, and speed of processing.