Image-to-Class Dynamic Time Warping And Its Application In3D Hand Gesture Recognition

3D Human Computer Interaction (HCI) becomes more and more popular thanks tothe emergence of commercial depth cameras. Moreover, hand gestures provide a naturaland attractive alternative to cumbersome interface devices for HCI. In this paper, wepresent an Image-to-Class Dynamic Time Warping (I2C-DTW) approach for3D handgesture recognition, which handles both static gesture and dynamic gesture.I2C-DTW classifier uses the descriptors distribute in a certain class and get bettergeneralization capability than using individual Image-to-Image measurements. Such aI2C warping path can be obtained by adding all the warping paths between thedescriptor distributions of the query sample and the certain class.The main idea of static gesture recognition is to divide the time-series curve of a3D hand gesture into various fnger combinations, called ‘fngerlets’, which canrepresent each gesture and capture inter-class variations. Furthermore, the I2C-DTWapproach searches for the minimal path to warp two fngerlets, which are from one testimage and the specifc class, respectively. Then the gesture recognition is to use theensemble of multiple image-to-class DTW distance of fngerlets to obtain betterperformance. We have evaluated this method on two datasets (10-Gesture, UESTC-ASL)and get99.5%and90.5%mean accuracy respectively.For dynamic hand gesture recognition, we extract the trajectory of hand andtransform it into two time-series curves, called gesture descriptor, which can representeach gesture trajectory and to capture inter-class variations. Moreover, the I2C-DTWapproach searches for the minimal path to warp two gesture descriptors, which are fromone test video clip and the specific class, respectively. Finally, the gesture trajectory isto use the ensemble of two image-to-class DTW distance of gesture descriptor to obtainbetter performance. We have evaluated this method on UESTC-DGL dataset and get98.44%mean accuracy. The experiment results show that the proposed I2C-DTWapproach signifcantly improves the recognizing performance.