Robust face tracking-by-detection via sparse representation

The robust algorithm, which is used for tracking human faces in unconstrained video, is built on Tracking-by-detection based on sparse representation. The algorithm works by combining the advantages of face tracking and face detection to improve the accuracy of tracking face in complex environment. The off-line trained face model fits input image to detect face and online trained tracker localizes face via sparse representation. Sparse representation makes human faces´ tracking more accurate and robust by the generalized Haar-like features. Also, it will make our tracking algorithm more adaptive and robust since it can be used for any original signal, which means K-sparse can be omitted. The algorithm is validated on a surveillance video considering complicated conditions, such as illumination variation, pose changes, and so on.