柴里弘毅 2 ,城 啓悟 3 ,加藤達也 2 Generally, non-contact gaze input devices require calibration before using. However, it is difficult Generally, non-contact gaze input devices require calibration before using. However, it is difficult for people with severe multiple disabilities who is mentally challenged and have difficulty in moving for people with severe multiple disabilities who is mentally challenged and have difficulty in moving their bodies as they wish, to move their eyes according to the

more »

... ions. It takes time and effort to their bodies as they wish, to move their eyes according to the instructions. It takes time and effort to use a dedicated device, which makes them difficult to express intentions by looking. In this study, we use a dedicated device, which makes them difficult to express intentions by looking. In this study, we constructed a real-time gaze area estimation system that does not require calibration while maintaining constructed a real-time gaze area estimation system that does not require calibration while maintaining the resolution required for challenged people. Using a usual web camera, gaze area estimation was the resolution required for challenged people. Using a usual web camera, gaze area estimation was realized by learning eye and facial appearance with Convolutional Neural Network (CNN) . Next, realized by learning eye and facial appearance with Convolutional Neural Network (CNN) . Next, 36 36 gazing points were set on the screen, and the evaluation experiments were performed by changing gazing points were set on the screen, and the evaluation experiments were performed by changing the relative distance between the camera and the face, and the posture angle of the face. In conclusion, the relative distance between the camera and the face, and the posture angle of the face. In conclusion, it was confirmed that practical accuracy of the basic posture was maintained up to 1,200 it was confirmed that practical accuracy of the basic posture was maintained up to 1,200 mm for the mm for the distance, 100 distance, 100 mm downward in terms of position. The results for the posture angle were obtained for one mm downward in terms of position. The results for the posture angle were obtained for one subject only, but it was shown that the accuracy was maintained up to 10 subject only, but it was shown that the accuracy was maintained up to 10 degrees for the yaw, 15 degrees for the yaw, 15 degrees degrees for the pitch, and 15 for the pitch, and 15 degrees for the roll angle, respectively. degrees for the roll angle, respectively. 一般に,非接触型の視線入力装置では,使用開始前にキャリブレーションを必要としている.しかし, 知的な遅れがあり身体を思うように動かすことが難しい重度重複障害者にとって,キャリブレーション の操作指示に従い視線を移動させることは困難である.専用の機器を必要とする手間がかかることも視 線による意思表示を難しくしている. 本研究では,意思表示に必要な分解能を維持しつつキャリブレーショ ンを必要としないリアルタイム視線領域推定システムを構築した.通常のwebカメラを用い,目や顔の情 報を畳み込みニューラルネットワーク (CNN)で学習することにより視線領域推定を実現している.次に, 画面上に36か所の注視点を設け,カメラと顔までの相対的な距離および顔の姿勢角を変化させて評価実 験を行った.その結果,基本姿勢において,距離については1,200 mm,位置では100 mm下方まで実用的 な精度が維持されていることが確認された.姿勢角については1人の被験者のみを対象とした結果である が,ヨー角10度,ピッチ角15度,ロール角15度までは精度が保たれることが確認された. (キーワード:視線領域推定,深層学習,重度重複障害,意思表示支援,アシスティブテクノロジー)