Abstract: The brain-computer Interface (BCI) gives interactive communications between people and the machine, and has fascinated the researchers over the last couple of years. However, the BCI system suffers from a low information transmission rate, low accuracy and poor interactive performance, which is the bottleneck for the promotion of BCI-actuated system. Therefore, to classify different motor commands fast with minimal error is an important problem in the BCI system. For the dynamic classification of motor imagery mind states in the brain-computer interface (BCI), we proposed a power projection based feature extraction method to classify the EEGs by combining information accumulative posterior Bayesian approach. This method improves the classification accuracy by maximizing the average projection energy difference of the two types of signals. The experimental results on two motor imagery datasets show that the maximum classification accuracy is about 90%. With three indexes, i.e. maximum classification accuracy, kappa coefficient and mutual information, the effectiveness of this method is demonstrated.