We propose to develop a video analysis system that detects suspicious, significant events using two methods: monitoring current activity against learned models of normal behavior, and searching for events explicitly indicated by the video analyst. The underlying behavior representations are based on object interactions in space and time, as well as temporal dynamics. Fusing anomaly detection and directed search should yield advantages of both. The system will be integrated with our VIRAT system, being developed on a major DARPA program, for content-based video retrieval based on actions and activities.
Keywords: Video Behavior Recognition, Video Analysis, Video Surveillance, Anomaly Detection, Activity Detection