モーションキャプチャで得られた顔特徴点位置データを用いた表情の識別

摘要

Our research goal is to build a dynamic deformation model of 3D face that creates natural facial expressions, through example-based learning of various facial expressions performed by human. As a preliminary step, we defined the feature vector representing 3D coordinates of the marker points attached on the face which were measured by means of a motion capture system, and discrimination capability of the feature among several categories of facial expressions was investigated by classification experiments. With a simple minimum distance classification method as well as dimensionality reduction of the feature vector achieved by PCA, humanlike ability or even better than human in discriminating facial expressions has been confirmed through evaluation testing of the classification accuracy conducted by leave-one-out method.%本研究ではモーションキャプチャで得られたマーカ点について，表情表出前の真顔に対応する位置座標との変位を特徴として求め，フレームごとに，この変位ベクトルを用いて最小距離識別法による表情の識別を行った場合の分類性能をLeave-One-Out法による識別実験で評価した．さらに，表情の違いを効率的に表す低次元の特徴ベクトルを主成分分析によって求め，これを用いて同様の識別実験を行い，結果の比較を行なった．