Head pose estimation is a crucial issue in human-robot-interaction (HRI) area. The existing methods are categorized in two approaches: geometry based and appearance based methods. However, geometry based methods are very weak to changes in large variation of pose, illumination, facial expression, and low resolution of the input image. Also, Appearance based methods have disadvantages on accuracy and continuity. In this paper, we propose a head orientation estimating method using deep convolutional neural networks (DNN). The proposed method is accurate, continuous, operating beyond real time, and robust to large variation of head pose and low resolution of an input image. Experimental result demonstrates our method outperforms the state-of-the-art method and very promising.