Human gesture recognition through a Kinect sensor

Abstract

Gesture recognition can be applied to many research areas, such as vision-based interface, communication and human robot interaction (HRI). This paper implements a non-intrusive, real-time gesture recognition system using a depth sensor. Related features are obtained from the human skeleton model generated by the Kinect sensor. Hidden Markov Models (HMMs) are used to model the dynamics of the gestures. We conducted offline experiments to check the accuracy and robustness of the system. Online experiments were also performed to verify the real-time requirement. Final results indicate that the average recognition accuracy is around 85% for the subject who provides the training data and 73% for the other subject who does not. The system was also used to interact with a mobile robot through gestures. This application indicates that it is robust to work in real-time.