Subways and other rapid transit systems are marked symbols of the modern metropolis. As a transdisciplinary service, accurately and safely positioning and tracking the metro trains helps the passengers to plan their travels and provides the operators with auxiliary information about the trains to enhance the metro system’s resilience. However, many general-purpose positioning technologies, such as Global Navigation Satellite Systems (GNSS) and Wi-Fi signals, do not apply to the situations of underground metro trains. In this paper, we propose a two-stage framework for automatic real-time tracking of metro cars implemented only with low-cost accelerometers, saving the expense for complicated infrastructures. In the off-line stage, reference maps are developed for station-to-station track sections using the onboard acceleration data. To handle the missing data and uncalibrated consumer-grade sensors, Gaussian process regression (GPR) is adopted to denoise and interpolate the online acceleration readings, followed by the application of the Kalman filter algorithm to track the cars in real-time with the help of the reference map. We tested the proposed system in Wuhan Metro Line 2, and the results showed that our system achieved an error below 5% in positioning.