The management and extraction of structured knowledge from large video recordings is at the core of urban/environment planning, resource optimization. We have addressed this issue for the networks of camera deployed in two underground systems in Italy. In this paper we show how meaningful events are detected directly from the streams of video. Later in an off-line analysis we can set this information into an adequate knowledge model representation that will allow us to model behavioral activity and obtain statistics on everyday people activities in metro station. Raw data as well as online and off-line metadata are stored in relational databases with spatiotemporal retrieval capabilities and allow the end-user to analyse different video recording periods.