Real time gesture recognition system using posture classifier and Jordan recurrent neural network
Yusuke Araga, Makoto Shirabayashi, Keishi Kaida, Hiroomi Hikawa · 2012
This paper proposes a Jordan recurrent neural network (JRNN) based dynamic hand gesture recognition system. A set of allowed gestures is modeled by a sequence of representative static images, i.e., postures. The proposed system first classifies the input postures contained in the input video frames, and the JRNN finds the input gesture by detecting the temporal behavior of the posture sequence. To enhance the ability of the JRNN to identify the temporal behavior of its input sequence, a new training method has been proposed. Due to the proposed method, the system can recognize the reverse gestures. Implemented on a PC equipped with a USB camera for the live image acquisition, the proposed system can process 12.5 frames per second (fps). Experimental results show that the system can recognize 5 gestures with the accuracy of 99.0%, and recognize 9 gestures with 94.3% accuracy, respectively.