Learning, detection and representation of multi-agent events in videos

摘要: In this paper, we model multi-agent events in terms of a temporally varying sequence sub-events, and propose novel approach for learning, detecting representing videos. The proposed has three main steps. First, order to learn the event structure from training videos, automatically encode sub-event dependency graph, which is learnt that depicts conditional between sub-events. Second, pose problem detection videos as clustering maximally correlated sub-events using normalized cuts. principal assumption made work are composed highly chain have high weights (association) within cluster relatively low (disassociation) clusters. does not require prior knowledge number agents involved an make any assumptions about length event. Third, recognize fact abstract should extend representations related human understanding events. Therefore, extension CASE representation natural languages allows plausible means interface users computer. We show results detection, meeting, surveillance, railroad monitoring domains.