Applications involving EEG pattern recognition are considered complicated, mainly due to the signal's non-stationary behavior. This characteristic negatively affects the development of BCI classification systems applied to rehabilitation devices, because signals acquired in equal conditions have differences between subjects and between multiple acquisitions of a single subject. Mainly, two approaches can be used to classify them: calibration and testing with signals acquired exclusively from the final user, or a single solution with greater generalization capabilities, which could be applied across different subjects. Developing in this sense, the present work presents an analysis of similarities between different subject EEG signals during the execution of a common task. The data was acquired from a competition dataset and was evaluated in relation to its similarities (using Dynamic Time Warping) and through classification using Common Spatial Patterns (CSP), Linear Discriminant Analysis (LDA), and Support Vector Machine (SVM). Four different experiments were performed, which showed most movement classes have similar distances across acquisition and across subjects. That was also seen in the classification, where SVM achieved the overall best results. The greatest accuracies occurred on training and testing with single subjects but simultaneously showed statistically similar results to experiments performed with cross-subject data, which indicates that the development of subject-independent solutions is possible.