For importance of urban railway transportation video surveillance and limitations of traditional video surveillance, a design using front-end image-processing module for dynamic background modeling, target detection and tracking was presented to model passenger behavior and identify abnormal passengers. Bayesian classifier was used to classify target and background for target detection, and prior knowledge of passenger characteristics was made into templates; with matching campaign area and target model for process track, passenger behavior changing in space and time was abstracted to data classification problems. This system was designed to satisfy railway transportation video surveillance requirement to ensure metro train running security.