The combination of Multiple-Input Multiple-Output (MIMO) signal processing with Orthogonal Frequency Division Multiplexing (OFDM) is regarded as a promising solution for enhancing the data rates of next-generation wireless communication systems operating in frequency-selective fading environments. However, when the length of MIMO OFDM symbols is larger than that of wireless channel delay, two intractable issues should be resolved before their applications in cellular fast fading channel scenarios with large numbers of users, i.e., the bandwidth overhead of channel estimation and the challenge to construct large numbers of orthogonal training sequences. So, a new design scheme of training sequence in time domain, is adopted to conduct channel estimation in MIMO OFDM systems, which works as a generations of Steiner method in multi-user CDMA uplink scenarios. Training sequences of different transmit antennas, can be simply obtained by truncating the circular extension of one basic training sequence and the pilot matrix assembled by these training sequences is one circular matrix with good reversibility. Furthermore, when the length of channel profiles is less than that of MIMO OFDM symbols, more bandwidth resources can be saved, as the training sequence only occupies a part of one MIMO OFDM symbol. Numerical results of bandwidth overheads and channel estimations, indicate the proposed method can save abundant bandwidth and achieve good channel estimation accuracy when compared with classical frequency and time domain approaches, respectively.