Acceleration feature extraction for handwritings with natural similarities

For natural and intuitive interaction, research on acceleration-based interfaces to recognize handwriting is becoming popular. When recognizing handwriting, a problem exists in distinguishing between naturally similar letters and numbers. This problem severely affects the total recognition rate. To solve this problem, Cho et al. used a two-stage recognizer, which separated confusing pairs from entire handwriting samples and then classified these during a second stage. In this paper, a feature extraction method to recognize similarly handwritten characters without using the two-stage procedure is investigated. As feature elements, temporal and statistical parameters from the accelerometer are calculated. Two feature extraction methods are proposed: creating vectors using these parameters, and remodeling each signal using these parameters. Dynamic Time Warping (DTW) is used to recognize these feature vectors. Results show a 3 dimensional feature vector is better at recognizing handwriting samples than a 2 dimensional vector, and the combination of acceleration and velocity is also better than other methods. From the overall results, a remodeled signal is acceptable in recognizing naturally similar handwritten characters.