Switches are critical infrastructure elements in the railway domain. Since they represent the nodes of the railway network, track changing is only possible at switches and their failure can have large impacts at the infrastructure availability. However, switches are exposed to harsh working conditions and prone to defects. In addition, their role as nodes of the railway network makes switches interesting for rail vehicle positioning. In this context, capturing vibration data from vehicle-borne sensors of in-service vehicles can serve two purposes: On the one hand, detecting switch passages and directions from the sensor data can help to resolve the path of the vehicle. On the other hand, the data of the detected switches can be used for switch condition monitoring. In this paper, we present a large experimental data set of switch passages collected with axle-box acceleration (ABA) sensors installed on a shunting locomotive operating on a mid-size port railway network. The switches are passed at various vehicle speeds and are of different types. The patterns of the different switches are analyzed in terms of repeatability of the patterns of the same switches and distinctiveness between switches. It is shown that classification into different switches as well as into trailing and facing passages is possible by means of neural network classifiers. Speed dependence on the patterns is investigated and modeled. Finally, applications for railway vehicle positioning and condition monitoring are demonstrated.