顔方向推定・開口検出に基づく複数人物のグループ化及び関係推定手法の検討

抄録

We are developing a system that groups pedestrians and identifies relation. After detecting multiple persons from input image, the system observes their faces and groups them based on estimation of distance and conversation. Conversation is detected from face direction and mouth opening. To track feature points in the small and low-resolution face image, we use an Active Appearance Model (AAM). In addition, the system needs to observe not only faces used in the training but also ones of others to deal with unspecified persons. To cope with change in face orientation, we prepare three types of face models (left / front / right views). To reduce the amount of training data of AAM, we consider that only one face is used as base face to create the model. The face parts of eye and nose extracted from the created face model are randomly moved and deformed to fit on the face of others. Thus, the model obtained by learning from one person is applied to others and the feature tracking gets robustness. The system judges whether they make a conversation with Support Vector Machine (SVM) using face direction and mouth opening. Experimental results show effectiveness of the proposed method.