In our every day life, our brain is constantly processing information and paying attention, reacting accordingly, to all sorts of sensory inputs (auditory, visual, etc.). In some cases, there is a need to accurately measure a person's level of attention to monitor a sportsman performance, to detect Attention Deficit Hyperactivity Disorder (ADHD) in children, to evaluate the effectiveness of neuro-feedback treatment, etc. In this paper we propose a novel approach to extract, select and learn spectral-spatial patterns from electroencephalogram (EEG) recordings. Our approach improves over prior-art methods that was, typically, only concerned with power of specific EEG rhythms from few individual channels. In this new approach, spectral-spatial features from multichannel EEG are extracted by a two filtering stages: a filter-bank (FB) and common spatial patterns (CSP) filters. The most important features are selected by a Mutual Information (MI) based feature selection procedure and then classified using Fisher linear discriminant (FLD). The outcome is a measure of the attention level. An experimental study was conducted with 5 healthy young male subjects with their EEG recorded in various attention and non-attention conditions (opened eyes, closed eyes, reading, counting, relaxing, etc.). EEGs were used to train and evaluate the model using 4x4fold cross-validation procedure. Results indicate that the new proposed approach outperforms the prior-art methods and can achieve up to 89.4% classification accuracy rate (with an average improvement of up to 16%). We demonstrate its application with a two-players attention-based racing car computer game.