This paper proposes a real-time method for recognizing finger character using range image. At first, we automatically extract the right hand region from range image which captured by Microsoft Kinect. Secondly we define sixe features of area ratio, aspect ratio, roundness, Foureir descriptor, number of figure, and range ratio, and these features are fed to SVMs. In order to prevent wrong recognition in a real-time process, we implement two functions of motionless judge and voting process. We set 41 Japanese finger characters without a motion as the target, and the evaluation experiment was carried out to one subject. As the result, we obtained the recognition rate of 82%. Moreover, there were 30 characters which acquired the recognition rate of 90% or more, and it was shown that our method is effective.%本論文では距離画像を用いたリアルタイムの指文字認識手法を提案する.Kinectにより得られた距離画像から右手領域を自動的に抽出する.次に手領域より面積比,アスペクト比,円形度,フーリエ記述子,指の本数および距離分布比の6種類の特徴量を求め,最後にSVMにより認識する.またリアルタイム処理において誤認識を防ぐために静止判定および投票処理を導入する.動きを伴わない41種類の日本指文字を認識対象とし,被験者1名に対して評価実験を実施した.その結果,認識率82％を得た.41種のうち90％以上の認識率を得た文字は30種あり,提案手法が有効であることを示した.
展开▼
