A major limitation of current Brain-Computer Interfaces (BCI) based on Motor Imagery (MI) is that they are subject-specific BCI, which require data recording and system training for each new user. This process is time consuming and inconvenient, especially for casual users or portable BCI with limited computational resources. In this paper, we explore the design of a Subject-Independent (SI) MI-based BCI, i.e., a BCI that can be used immediately by any new user without training the BCI with the user's data. This is achieved by training the BCI on data acquired from several other subjects. In order to assess the possibility to build such a BCI, we compared several designs based on different features and classifiers, on data from 9 subjects. Our results suggested that linear classifiers were the most appropriate for the design of MI-based SI-BCI. We also proposed a filter bank common spatial patterns feature extraction method based on a multi-resolution frequency decomposition which achieved the highest accuracy.