Improved Face Tracking Thanks to Local Features Correspondence

In this paper, we propose a technique to enhance the quality of detected face tracks in videos. In particular, we present a tracking algorithm that can improve the temporal localization of the tracks, remedying to the unavoidable failures of the face detection algorithms. Local features are extracted and tracked to “fill the gaps” left by missed detections. The principal aim of this work is to provide robust and well localized tracks of faces to a system of Interactive Movietelling, but the concepts can be extended whenever there is the necessity to localize the presence of a determined face even in environments where the face detection is, for any reason, difficult. We test the effectiveness of the proposed algorithm in terms of faces localization both in space and time, first assessing the performance in an ad-hoc simulation scenario and then showing output examples of some real-world video sequences.