Abstract: To cope with the anormalities within the throat of a marshalling station, such as the intrusion of foreign objects and the left of tools and instruments, an intelligent visual anomaly detection system based on unsupervised learning is studied and implemented. The 5th generation mobile communication technology(5G) slicing technology is used to transimmit 4K ultra-high definition video data collected from the throat and the operator's User Plane Function(UPF) sinks to keep the data not outbound so as to ensure network security. In view of the uncertainty of the anomalities, the student-teacher feature pyramid matching algorith, which is an unsupervised learning method, is adopted to recognize the abnormalities within the throat by comparing the score of the teacher network and that of the student network. The system has achieved good results in the on-site test at Huaihua West marshalling station and can effectively improve the safety prevention ability of marshalling station throat.