Multi-clue based multi-camera face capturing

This paper presents a novel face capturing approach that integrates several cooperative cameras using multiple clues. Being different from face detection, face capturing means active face gathering by cameras. Our idea is that face pose can be separated into adjacent pose subspaces by cooperative cameras in different directions. In this way, face capturing may be easier than face detection since face detection probably surfers from the pose problem if people do no like to see cameras directly. When different cameras are mounted in different face directions, high performance face detection algorithms can be used in each camera. Consequently, the multi-pose face capturing problem is transformed to face correspondence among cameras after face detection. In the proposed approach, face detection is first carried out in each camera. The detected faces are then corresponded by dynamic programming. For efficient face correspondence, the object features (face position, face appearance) and context features (clothes color, face order) are elaborately combined in the feature extraction and feature matching step. Experiments are carried out in a real scene. The results showed very good performance with 95.6% correct correspondence and 0.9% errors.