Over the last few decades, human action recognition has become one of the most challenging tasks in the field of computer vision. Employing economical depth sensors such as Microsoft Kinect as well as recent successes of deep learning approaches in image understanding has led to effortless and accurate extraction of 3D skeleton information. In this study, we have introduced a novel bag-of-poses framework for action recognition by exploiting 3D skeleton data. Our assumption is that any action can be represented with a set of predefined spatiotemporal poses. The pose de-scriptor is composed of two parts, the first part is concatena-tion of the normalized coordinate of the skeleton joints. The second part consists of temporal displacement of the joints which is constructed with predefined temporal offset. In order to generate the key poses, we apply K-means clustering overall training pose descriptors of dataset. To classify an action pose, we train a SVM classifier with the generated key poses. Thereby, every action on dataset is encoded with key-poses histogram. We use ELM classifier to recognize the actions since it has been shown to be faster, accurate , and more reliable than other classifiers. The proposed framework is validated with four publicly available benchmark 3D action datasets. The results show that our frame-2 Saeid Agahian et al. work achieves state-of-the-art results on three of the datasets compared to the other methods and produces competitive result on the fourth.