Using Video Surveillance to Detect Dangerous Situations in Underground Stations by Computer Vision

Underground stations traditionally are a common application for video surveillance. Currently, humans have to be employed to control trains and stations to avoid harm to passengers and equipment using closed-circuit television (CCTV), but this personell has to watch at least 20 screens at the same time and could miss important events. However, the already installed surveillance equipment in stations can be combined with algorithms presented in this paper to let a computer raise an alarm when certain conditions are met, typically caused by passengers walking on the tracks, throwing objects or being stuck in doors while the train is moving. People that are too close to the tracks while a train enters the station should cause a proximity warning. In order to detect these events, multiple cameras have to be connected to a single processing unit, and certain areas of interest have to be defined beforehand. Since no markers are carried by passengers, natural feature tracking has to be implemented, which is used to classify every person to belong to a certain area. Further, big crowds can lead to mass hysteria and have to be detected in time to let emergency personell prepare to intervene. All these techniques have analyze the data in real time. Since false alarms can cause significant delays in regular operations, the detection has to be fairly accurate and tolerant to environment changes.∗