Light-weighted Network based Human Pose Estimation for Mobile AR Service

Recent advances in deep learning algorithms and hardware have led to the rapid development of computer vision. Based on the development of computer vision field, many deep learning algorithms are applied to augmented reality(AR) technology. However, in order to increase the accessibility of users, the majority of AR services are provided in mobile devices. In addition, many deep learning algorithm techniques require high computing resources to be applied directly to AR services. Therefore, in order to apply deep learning technology to mobile AR services, it is necessary to implement a lightweight network. In this paper, we propose a light-weighted network model to estimate the single human pose for the mobile AR service. The contribution of this paper is as following. i) The proposed light-weighted network is applied to the commercial mobile devices. ii) A method for integrating Unity 3D rendering tool and TensorFlow Lite library is proposed. iii) The body skeleton of human is extracted and analyzed in real time. Our results suggest that it will be useful to provide a way for new interactions in mobile AR services.