Dynamic Time Warping-Based K-Means Clustering for Accelerometer-Based Handwriting Recognition

Dynamic time warping(DTW) is widely used for accelero-meter-based gesture recognition. The basic learning strategy applied with DTW in most cases is instance-based learning, where all the feature vectors extracted from labeled training patterns are stored as reference patterns for pattern matching. With the brute-force instance-based learning, the number of reference patterns for a class increases easily to a big number. A smart strategy for generating a small number of good reference patterns is needed. We propose to use DTW-based K-Means clustering algorithm for the purpose. Initial training is performed by brute-force instance-based learning, and then we apply the clustering algorithm over the reference patterns per class so that each class is represented by 5 ~ 10 reference patterns each of which corresponds to the cluster centroid. Experiments were performed on 5200 sample patterns of 26 English uppercase alphabets collected from 40 personals using a handheld device having a 3-d accelerometer inside. Results showed that reducing the number of reference patterns by more than 90% decreased the recognition rate only by 5%, while obtaining more than 10-times faster classification speed.