Most conventional semi-blind channel estimation schemes for MU-MIMO systems are based on eigenval decomposition (EVD) or singular value decomposition (SVD). However, EVD-or SVD-based channel estimation would impose a high computational complexity when the base station is equipped with large number antennas. Those methods are not well suited for real-time processing, especially for Channel State Information (CSI) tracking in time-varying environments. In order to reduce the computational complexity, a PASTd-based (Projection Approximation Subspace Tracking with deflation) CSI tracking algorithm is proposed. The PASTd-based algorithm converges fast, has low computational complexity, and can operate with very small training overhead. Simulation results show that the proposed algorithm can effectively track the CSI with mild Doppler spreads. Only one pilot symbol per user at the beginning of transmission session is needed to resolve the multiplicative factor ambiguity.