DTW Based Clustering to Improve Hand Gesture Recognition

Vision based hand gesture recognition systems track the hands and extract their spatial trajectory and shape information, which are then classified with machine learning methods. In this work, we propose a dynamic time warping (DTW) based pre-clustering technique to significantly improve hand gesture recognition accuracy of various graphical models used in the human computer interaction (HCI) literature. A dataset of 1200 samples consisting of the ten digits written in the air by 12 people is used to show the efficiency of the method. Hidden Markov model (HMM), input-output HMM (IOHMM), hidden conditional random field (HCRF) and explicit duration model (EDM), which is a type of hidden semi Markov model (HSMM) are trained on the raw dataset and the clustered dataset. Optimal model complexities and recognition accuracies of each model for both cases are compared. Experiments show that the recognition rates undergo substantial improvement, reaching perfect accuracy for most of the models, and the optimal model complexities are significantly reduced.