Discriminating EEG signals between different motor imagery states is an important application of brain computer interface (BCI). However, low signal-to-noise ratio and significant data variation of EEG make it very difficult for BCI to obtain reliable results. Spatial filtering is one of the most successful feature extraction methods, and many efforts have been made to construct spatial filters that are robust against the data nonstationarity. In this paper, we propose a novel spatial filter optimization method based on mutual information which is estimated by Gaussian functions instead of Parzen window. We analyze the relationship between the mutual information and feature distance using a simulation study to show that optimization based on mutual information can contribute to feature stationarity. Moreover, we also evaluate the proposed method on a real world motor imagery EEG data set recorded from 16 subjects performing motor imagery or staying in idle state. The experimental results validate the effectiveness of the proposed spatial filter optimization method as it outperforms both the common spatial pattern analysis and filter-bank common spatial pattern analysis.