2D CNN and Gated Recurrent Network for Dynamic Hand Gesture Recognition with A Fusion of RGB-D and Optical Flow Data
Sunil A. Patel1, Ramji M. Makwana2
1Sunil A. Patel, Gujarat Technological University, Ahmedabad, Gujarat, India, Computer Engineering Department.
2Dr. Ramji M. Makwana, Managing Director, AIIVINE PXL Pvt. Ltd, Rajkot, Gujarat, India
Manuscript received on 02 July 2019 | Revised Manuscript received on 09 July 2019 | Manuscript published on 30 August 2019 | PP: 1784-1792 | Volume-8 Issue-10, August 2019 | Retrieval Number: J91850881019/2019©BEIESP | DOI: 10.35940/ijitee.J9185.0881019
Open Access | Ethics and Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: The dynamic hand gesture is an essential and important research topic in human-computer interaction. Recently, Deep convolutional neural network gives excellent performance in this area and gets promising results. But the Researcher had focused less attention on the feature extraction process, unification of frame, various fusion scheme and sequence-to-sequence prediction of a frame. Therefore, in this paper, we have presented an effective 2D CNN architecture with three stream networks and advances weighted feature fusion scheme with the gated recurrent network for dynamic hand gesture recognition. To obtain enough and useful information we have converted each RGB-D video to 30-frame and 45-frame for input. We have calculated an optical flow for frame-to-frame by given RGB video and extract dense motion features. After finding proper motion path, we have assigned more weight to optical flow features and fuse this information to the next stage and gets a comparable result. We have also added a newest Gated recurrent network for temporal recognition of frame and minimize training time with improved accuracy. Our proposed architecture gives 85% accuracy on the standard VIVA dataset.
Keywords: 2D Convolutional neural network, Gesture recognition, Optical flow, RGB-D data, Gated recurrent unit, weighted fusion
Scope of the Article: Optical and High-Speed Access Networks