|
Real-time face recognition from uncontrolled video sequences is adifficult problem. The difficulties come from large variation inface size, position, pose, expression, lighting condition andbackground. Also, the real-time constraint makes this problem moretough due to limited computation resources. In this thesis, we propose a fully automatic procedure that candetect faces, estimate the size of a face, segment the facefrom the scene by utilizing the knowledge about human motion.The searching operation for detecting faces is essentiallyone-dimension-based and thus is very efficient. The detected faceis tracked using the temporal consensus of head motion.During tracking the face, the system extracts features frommultiple views of the face regions by KLT. The extractedfeatures are clustered using a locally unsupervised and globallysupervised learning algorithm. We demonstrate that this systemcan recognize faces in real time with significant variation in headsize, facial expression, lighting condition, and background.Also, by introducing a rejection threshold for badly extracted features,the system can recognize all (100%) of the 88 persons in our database
|