In this paper, a comprehensive scheme is proposed for unconstrained joint face detection and recognition in video sequences for surveillance systems. Unlike conventional video based face recognition techniques, emphasis is laid on the acquisition of a pose constrained training video database followed by the extraction of well aligned face images from the training videos. We have proposed a new Indian Faces Video Database (IFVD) to demonstrate the performance of the proposed approach especially in the challenging environment of varying skin color and texture of faces from the Indian subcontinent.Our approach produces successful face tracking results on over 86% of all videos. The good tracking performance induces high recognition rates: 85.86 on Honda/UCSD and over 77.49 % on IFVD. The proposed technique is robust and aims to develop a unified framework to address the challenges of varying head orientation, pose and illumination level in a highly integrated fashion so as to benefit from the interdependence between the high fidelity face detection and the subsequent recognition phases.
展开▼
