This paper aims at developing a proper acceleration estimate of a train using a low-cost MEMS IMU (Inertial Measurement Unit) equipped with a 3-axis accelerometer and a 3-axis gyrometer. In particular, the impact of adding a velocity observation on the convergence of the method is studied. Compared to conventional methods that use the most often velocity sensors such as GPS or tachymeter, the proposed approach estimates the train velocity from a spectral analysis of the z-axis accelerometer only. This observation is then integrated in the measurement vector of a factorized Kalman filter in addition to the measurements of accelerations and angular rates as well as an orientation estimate based on acceleration measurements. Moreover, a realtime setting applied on the process and measurement noise matrices is introduced allowing an adaptive Kalman filtering. At last, experimental results performed on a urban train are presented and detailed.