Face recognition from a moving platform via sparse representation

Abstract

A video-based surveillance system for passengers includes face detection, face tracking and face recognition. In general, the final recognition result of the video-based surveillance system is usually determined by the cumulative recognition results. Under this strategy, the correctness of face tracking plays an important role for the system recognition rate. For face tracking, the challenges of face tracking on a moving platform are that the space and time information used for conventional face tracking algorithms may be lost. Consequently, conventional face tracking algorithms can barely handle the face tracking on a moving platform. In this paper, we have verified the state-of-the-art technologies for face detection, face tracking and face recognition on a moving platform. In the mean time, we also proposed a new strategy for face tracking on a moving platform or face tracking under very low frame rate. The steps of the new strategy for face detection are: (1) classification the detected faces over a certain period instead of every frame (2) Tracking of each passenger is equivalent to reconstruct the time order of certain period for each passenger. If the cumulative recognition results are the only part needed for the surveillance system, step 2 can be skipped. In addition, if the additional information from the passengers is required, such as path tracking, lip read, gesture recognition, etc, time order reconstruction in step 2 can offer the information required.