In this article we present a real time platform for semantic video interpretation applied to bank agency monitoring. The proposed system is a multi-camera platform, which recognizes userpredefined scenarios, such as bank attack scenarios. These scenarios are modelled by domain experts using a back and forth process and based on a representation language. In order to address bank monitoring issues, the system has been improved based on two evaluation types. First, a repair stage guided by a careful technical evaluation has been performed at each level of the interpretation chain. As a consequence, the robustness obtained was sufficient enough to recognize all scenarios of interest. Second, an end-user evaluation has helped the experts to improve the scenario models to adapt them to real life situations. We report results of scenario recognition performances on real video sequences taken in a bank agency.

No abstract