Online Kinect Handwritten Digit Recognition Based on Dynamic Time Warping and Support Vector Machine

Abstract

Handwriting in-space from Kinect depth and color information is a challenging task due to the high variability of signature characteristics for difierent individuals. In this paper, a user-friendly human computer interaction system is proposed and implemented based on Kinect handwriting. The flngertip is flrstly tracked by our detection method in every depth frame to generate 3D trajectory of handwriting, and then normalization and smoothing are performed before feature extraction. On this basis, the time sequence feature of 3D signature can be captured as an online character recognition method, and a joint recognition framework is proposed based on DTW and SVM for input vectors of difierent lengths. The evaluation on a handwriting in-space dataset of digits from 0 to 9 shows that the proposed recognition scheme can ofier a high recognition accuracy and a satisfying robustness to noisy data in digit recognition test even with small training number. Therefore, the method can be successfully applied in many Human Computer Interaction applications in real world.