The development of a scalable and adaptable trackside and on-board coordinated control system (CCS+) is a top priority for the digital and automated train operations of future railways. One of the most important and reliable solutions in CCS+ is train positioning, which is called GLAT (Generic Location-Awareness Tools). The technologies of GLAT can rely on many methods like GNSS such as INS/GPS. In this paper we introduce a deep learning method to recognize the location through only monocular cameras. This method can be very useful when, for example, in "satellite signal blind spots" such as tunnels and mountainous areas or even no GPS signals at all. We carried out the first visual train positioning research experiment in the China National Railway Test Center (East Suburban Branch). By using a well-trained Train Localization Network, we demonstrate that trains can successfully recognize the real-time location when it moves along its route. Compared with traditional INS/GNSS integrated navigation scheme, the computer vision method is more robust and attractive since it brings trains an "intelligent brain", the train can deduce the localization by exploring its own "visual memory". This kind of capability can accelerate more intelligence utilized by modern trains.