Abstract/Summary

This paper addresses the complex problem of recognising threat situations from videos streamed by surveillance cameras. A behaviour recognition approach is proposed, which is based on a semantic recognition of the event. Low-level tracking information is transformed into high-level semantic descriptions mainly by analysis of the tracked object speed and direction. Semantic terms combined with automatically learned activity zones of the observed scene allow delivering behaviour events indicating the mobile activity. Behaviours of interest are modelled and recognised in the semantic domain. The approach has been applied on different public datasets, namely CAVIAR and ARENA. Both datasets contain instances of people attacked (with physical aggression). Successful results have been obtained when compared to other state of the art algorithms.