Emotions play a significant role in human communication and decision making. In order to bypass current limitations of human-robot interaction, more natural, trustworthy and nonverbal way of communication is needed. This requires robots to be able to explain and perceive person's emotions. Our work is based on the concept that each emotional state can be placed on a two-dimensional plane with arousal and valence as the axes. We propose a new feature set based on using the linear predictive coefficients on wavelet-decomposed EEG signals. Emotion classification is then performed using support vector machine with Gaussian kernel. Proposed approach is evaluated on EEG signals from publicly available DEAP dataset and results show that our method is effective and outperforms some state of the art methods.