Gesture imitation using machine learning techniques

This study is a part of an ongoing project which aims to assist in teaching Sign Language (SL) to hearing-impaired children by means of non-verbal communication and imitation-based interaction games between a humanoid robot and a child. In this paper, the problem is geared towards a robot learning to imitate basic upper torso gestures (SL signs) using different machine learning techniques. RGBD sensor (Microsoft Kinect) is employed to track the skeletal model of humans and create a training set. A novel method called Decision Based Rule is proposed. Additionally, linear regression models are compared to find which learning technique has a higher accuracy on gesture prediction. The learning technique with the highest accuracy is then used to simulate an imitation system where the Nao robot imitates these learned gestures as observed by the users.