Activity Recognition and Room-Level Tracking in an Office Environment

We present an approach for multi-person activity recognition in an office environment with simultaneous tracking of users on the room-level. Audio as well as video features, gathered from a simple setup, are used to employ a multilevel hidden Markov model (HMM) framework. Evaluation on unconstrained real world data recorded on several days in five offices with one camera and one microphone per room is presented for activity recognition. We track the users by a distributed camera network which has to cope with blind gaps between different camera views. For location estimation, we apply a Bayesian filter on top of the activity recognition results. Results on a dedicated tracking sequence of one hour length show the algorithm's performance