This paper proposes a novel approach to pose-based human action recognition. Given a set of training images, we first extract a scale invariant contour-based pose feature from silhouettes. Then, we cluster the features in order to build a set of prototypical key poses. Based on their relative discriminative power for action recognition, we learn weights that favor distinctive key poses. Finally, classification of a novel action sequence is based on a simple and efficient weighted voting scheme that augments results with a confidence value which indicates recognition uncertainty. Our approach does not require temporal information and is applicable for action recognition from videos or still images. It is efficient and delivers real-time performance. In experimental evaluations for single-view action recognition and the multi-view MuHAVi data set, it shows high recognition accuracy.