Detecting Single-Actor Events in Video Streams for TRECVid 2008

This paper presents the systems and results for the Event Detection task of the TRECVid 2008 evaluation campaign. The kind of events addressed are single-actor, without requiring detailed tracking of body parts; rather the body as a whole is considered and the interaction between people is not addressed. The events thus considered are the OpposingFlow, PersonRuns and ElevatorNoEntry. For the first two, motion vectors of areas with large frame-by-frame activity are analyzed. Groups of motion vectors pointing opposite at key door-frames indicate opposing flow events, while other groups with large magnitudes that persist across frames indicate running. For the ElevatorNoEntry event, the state of the elevator doors and the number of people present in the camera field of view are tracked to decide if an elevator arrives and none of the people present enter it. People and elevator doors are tracked using blob-based tracking on top of adaptive foreground segmentation.

• Type: Conference
• Year: 2008
• Where: TRECVID
• Authors: Andreas Stergiou, Aristodemos Pnevmatikakis, Lazaros Polymenakos, Nikos Katsarakis
• Keywords: Elevator Doors, Event Detection Task, Internet Technology, Motion Vectors, TRECVID 2008