Spatio-temporal Keypoints for Video-Based Face Recognition

A new approach for video-based face recognition is presented in this paper. The proposed technique is based on the use of key points and related descriptors for image representation. In particular this work introduces a general approach to extend to the temporal dimension the analysis usually carried out on single images, with the aim of deriving a more stable and compact representation of video information. The algorithm, which can easily be coupled with different techniques for key point extraction and representation, has been designed taking into account two fundamental requirements in this application scenario: recognition accuracy and efficiency of the matching process. The extensive experiments carried out on public datasets using a specific implementation based on SURF demonstrate the validity of the proposed technique.