Human Action Recognition using Convolutional Neural Network: Case of Service Robot Interaction

This paper proposes a Human Robot Interaction (HRI) framework for a service robot capable of understanding common interactive human activities. The human activity recognition (HAR) algorithm is based on end to end deep Convolutional Neutral Network architecture. It uses as an input a view invariant 3D data of the skeleton joints, which is recorded from a single Microsoft Kinect camera to create a specific dataset of six interactive activities. In addition, an analysis of the most informative joint is made in order to optimize the recognition process. The system framework is built on Robot Operating System (ROS), and the real-life activity interaction between our service robot and the user is conducted for demonstrating the effectiveness of the developed HRI system. The trained model is evaluated on an experimental dataset created for this work and also the publicly available datasets Cornell Activity Dataset (CAD-60), and KARD HAR datasets. The performance of the proposed algorithm is proved when compared to other approaches and the results confirm its efficiency.