In this study, we aim to develop a robust motion recognition system for an intelligent video surveillance system, that can be used for security, sports and rehabilitation by using extended alternative learning. A robust motion recognition system is necessary for the automated detection of security incidents by using a machine learning approach. However, to avoid the difficulty of collecting a huge training dataset, we propose an alternative learning approach that trains a deep neural network with a 3D-CG dataset to recognize several motions. We present our experimental results on motion recognition from free-viewpoint videos by using deep learning and alternative learning. The trained deep neural network (DNN) is evaluated using actual videos by classifying the different actions performed by real humans in these videos.