Costume: A New Feature for Automatic Video Content Indexing

This paper deals with the introduction of costume as a new feature for automatic video content indexing. We present in this paper an application of person recognition using costumes, in order to show the relevance of costume for indexation. The recognition is carried out by extracting the costume of all the persons who appear in the video. Then, their reappearance in subsequent frames is performed by searching the reappearance of their costume. The human presence is detected by searching faces, and then a feature for costume is extracted according to the scale and the location of each face. The Bhattacharyya coefficient, which is a coefficient derived from the Bayes error, is used to compare the color distribution of the various costumes. Finally, some results of people recognition are presented, as well as different axes for further research.