Multi-source Multi-modal Activity Recognition in Aerial Video Surveillance

摘要: Recognizing activities in wide aerial/overhead imagery remains a challenging problem due part to low-resolution video and cluttered scenes with large number of moving objects. In the context this research, we deal two un-synchronized data sources collected real-world operating scenarios: full-motion videos (FMV) analyst call-outs (ACO) form chat messages (voice-to-text) made by human watching streamed FMV from an aerial platform. We present multi-source multi-modal activity/event recognition system for surveillance applications, consisting of: (1) detecting tracking multiple dynamic targets platform, (2) representing target tracks as graphs attributes, (3) associating using probabilistic graph-based matching approach, (4) spatial-temporal activity boundaries. also pattern learning framework which uses associated training index archive videos. Finally, describe multi-intelligence user interface querying interest (AOIs) movement type geo-location, playing-back summary text segments targets-of-interest (TOIs) (in both pixel geo-coordinates). Such tools help end-user quickly search, browse, prepare mission reports data.