Mouth Tracking from Video Sequences using Trainable Multivariate Gaussian Classifiers

概要

論文の詳細を見る
In this work, we describe trainable multivariate Gaussian mixture classifiers for mouth tracking from video sequences. The proposed algorithm is developed within the Bayesian framework. The mouth tracking is a crutial task for large class of applications such as audio-visual speech recognition, and talking humanoid head design. The method uses a learned image representation of pixel region characteristics, based upon the color and derived texture features. Both positive and negative exemplars of some visually apparent characteristic which forms the basis of the seperation, input by the user using a graphical user interface system, are used together with a clustering algorithm to construct positive similarity and negative similarity multivariate Gaussian mixture classifiers. Classification is accomplished via these classifiers, whereby linguistic meaning is assigned to each pixel in the image set. The largest connected blob with geometric information is segmented as the mouth region while the other smaller patches are regarded as outliers. Then, the amne parameters are removed from the segmented mouth image for consistent visual information processing.
社団法人電子情報通信学会の論文
2003-12-12