Low cost RGB-D sensors have been used extensively in the field of Human Action Recognition. The availability of skeleton joints simplifies the process of feature extraction from depth or RGB frames, and this feature fostered the development of activity recognition algorithms using skeletons as input data. This work evaluates the performance of a skeleton-based algorithm for Human Action Recognition on a large-scale dataset. The algorithm exploits the bag of key poses method, where a sequence of skeleton features is represented as a set of key poses. A temporal pyramid is adopted to model the temporal structure of the key poses, represented using histograms. Finally, a multi-class SVM performs the classification task, obtaining promising results on the large-scale NTU RGB+D dataset.