A hybrid unsupervised/supervised model for group activity recognition

Abstract

The new method proposed here recognizes activities performed by a group of users (e.g., attending a meeting, playing sports, and participating in a party) by using sensor data obtained from the users. Note that such group activities (GAs) have characteristics that differ from those of single user activities. For example, the number of users who participate in a GA is different for each activity. The number of meeting participants, for instance, may sometimes be different for each meeting. Also, a user may play different roles (e.g., `moderator' and `presenter' roles) in meetings on different days. We introduce the notion of role into our GA recognition model and try to capture the intrinsic characteristics of GAs with a hybrid unsupervised/supervised approach.