Human Robot Interaction Using Dynamic Hand Gestures

Abstract

This paper describes the implementation of a robust dynamic hand gesture recognizer using a depth sensor. The recognizer uses only depth image information, and the hand position provided by a hand tracker library, in order to construct its feature vectors. The recognizer builds two types of feature vectors to increase accuracy; the frame feature vectors that describe a static hand, and the sequence feature vectors that describe a contiguous segment of frames. The recognizer also uses two statistical classifiers. The frame feature vectors are utilized by the frame classifier. The results of the classifier, then become part of the sequence feature vector, which in turn are utilized by the sequence classifier. The results show that the accuracy of the recognizer increases more than twice, when using both classifiers. The recognizer also does not make any assumption for when a gesture begins or when it ends. Instead it learns to differentiate between noise, and a real gesture. A humanoid robot, ROBIN, is used for validation of the approach for human-robot interaction.