A new method of detecting out-of-round wheels from in-service vehicles is developed using characteristics of cycle impact between out-of-round wheel and track. The initial moment and the end moment as a rotation of wheel, which are determined through integration of the operational velocity of train, are used to truncate and divide the time history of vertical axle accelerations. The truncated acceleration signal between initial moment and the end moment was transformed by frequency sliced wavelet transform (FSWT) to get the time-frequency characteristic figures of wheel-rail impact. The time axis was replaced by displacement axis signifying wheel circumference in the time-frequency figures to achieve time-frequency feature locations as a rotation of wheel. The adjacent time-frequency feature locations during different rotation of wheel were compared to detect out-of-round wheels. The effectiveness of way has been proved through the dynamics simulation data, with high real-time, high accuracy and a good engineering applicability.