A preliminary study of micro-gestures : dataset collection and analysis with multi-modal dynamic networks
1University of Oulu, Faculty of Information Technology and Electrical Engineering, Department of Computer Science and Engineering, Computer Science and Engineering
|Online Access:||PDF Full Text (PDF, 2.5 MB)|
|Persistent link:|| http://urn.fi/URN:NBN:fi:oulu-201706022490
|Publish Date:|| 2017-06-02
|Thesis type:||Master's thesis
Micro-gestures (MG) are gestures that people performed spontaneously during communication situations. A preliminary exploration of Micro-Gesture is made in this thesis. By collecting recorded sequences of body gestures in a spontaneous state during games, a MG dataset is built through Kinect V2. A novel term ‘micro-gesture’ is proposed by analyzing the properties of MG dataset. Implementations of two sets of neural network architectures are achieved for micro-gestures segmentation and recognition task, which are the DBN-HMM model and the 3DCNN-HMM model for skeleton data and RGB-D data respectively. We also explore a method for extracting neutral states used in the HMM structure by detecting the activity level of the gesture sequences. The method is simple to derive and implement, and proved to be effective. The DBN-HMM and 3DCNN-HMM architectures are evaluated on MG dataset and optimized for the properties of micro-gestures. Experimental results show that we are able to achieve micro-gesture segmentation and recognition with satisfied accuracy with these two models. The work we have done about the micro-gestures in this thesis also explores a new research path for gesture recognition. Therefore, we believe that our work could be widely used as a baseline for future research on micro-gestures.
© Chen Haoyu, 2017. This publication is copyrighted. You may download, display and print it for your own personal use. Commercial use is prohibited.