In this paper, we propose an approach for recognizing human actions based on motion sequence information in RGB-D video using deep learning. A new representation that gives emphasis to the key poses associated with each action is presented. The features obtained from motion in RGB and depth video streams are given as input to the convolutional neural network to learn the discriminative features. The efficacy of the proposed approach is demonstrated on MIVIA action, NATOPS gesture, SBU Kinect interaction, and Weizmann datasets. © 2017 Elsevier Ltd