A novel competitive neural classifier for gesture recognition with small training sets

Abstract

Gesture recognition is a major area of interest in human-computer interaction. Recent advances in sensor technology and Computer power has allowed us to perform real-time joint tracking with com-modity hardware, but robust, adaptable, user-independent usable hand gesture classification remains an open problem. Since it is desirable that users can record their own gestures to expand their gesture vocabulary, a method that performs well on small training sets is required. We propose a novel competitive neural classifier (CNC) that recognizes arabic numbers hand gestures with a 98% success rate, even when trained with a small sample set (3 gestures per class). The approach uses the direction of movement between gesture sampling points as features and is time, scale and translation invariant. By using a technique borrowed from ob-ject and speaker recognition methods, it is also starting-point invariant, a new property we define for closed gestures. We found its performance to be on par with standard classifiers for temporal pattern recognition.