3-D hand trajectory recognition for signing exact english

Abstract

This paper presents a hierarchical approach to recognize isolated 3-D hand gesture trajectories for Signing Exact English (SEE). SEE hand gestures can be periodic as well as non-periodic. We first differentiate between periodic and non-periodic gestures followed by recognition of individual gestures. After periodicity detection, non-periodic trajectories are classified into 8 classes and periodic trajectories are classified into 4 classes. A Polhemus tracker is used to provide the input data. Periodicity detection is based on Fourier analysis and hand trajectories are recognized by Vector Quantization Principal Component Analysis (VQPCA). The average periodicity detection accuracy is 95.9%. The average recognition rates with VQPCA for non-periodic and periodic gestures are 97.3% and 97.0% respectively. In comparison, k-means clustering yielded 87.0% and 85.1%, respectively.