Face detection and clustering for video indexing applications

This paper describes a method for automatically detecting human faces in generic video sequences. We employ an iterative algorithm in order to give a confidence measure for the presence or absence of faces within video shots. Skin colour filtering is carried out on a selected number of frames per video shot, followed by the application of shape and size heuristics. Finally, the remaining candidate regions are normalized and projected into an eigenspace, the reconstruction error being the measure of confidence for presence/absence of face. Following this, the confidence score for the entire video shot is calculated. In order to cluster extracted faces into a set of face classes, we employ an incremental procedure using a PCA-based dissimilarity measure in con-junction with spatio-temporal correlation. Experiments were carried out on a representative broadcast news test corpus.