A superimposed-training (ST) based rapidly time-varying frequency selective channel estimation scheme for orthogonal frequency division multiplexing (OFDM) system is investigated. The time-varying coefficients of the rapidly time-varying channel are modeled by exponential basis expansion. Meanwhile, file optimal training sequences are derived with respect to minimizing the mean square error (MSE). Furthermore, an iterative technique for symbol detector is proposed to cancel the effects of the embedded training on symbol recovery. Simulation results confirm the effectiveness of the proposed approach, even for the rapidly time-varying channel with a great Doppler frequency.