Self co-articulation removal and hybrid classifier-feature combination for dynamic hand gesture recognition

Abstract

Gestures have been considered as a way to reinforce information as interfaces that can effectively deliver more natural, creative and intuitive methods for communication with devices. During gesticulation, some unwanted strokes known as self co-articulated strokes occur which create confusion between many of the gestures. Existing approaches track the gestures with single strokes but fails to remove the unwanted strokes. In order to recognize the correct gestures, these self co-articulated strokes need to be identified and removed from the gesture trajectory. In this paper, a hand gesture recognition system is proposed for bare-hand dataset which tries to solve the problem of self co-articulation of gestures present in numerals and alphabets. A two-step slope-angle system is used for detecting and removing the self co-articulated strokes. These strokes which are removed from the tracked trajectory can be used as added features along with proposed features. Twelve new features like curliness feature, sharpness feature etc. are being extracted and added to existing features to form a feature matrix of forty features. Recognition accuracy for different classifiers such as SVM, k-NN, ANN, ELM, Naïve Bayes and Decision tree along with hybrid classifier–feature combination models is being calculated. The experimental results suggest that highest accuracy achieved for individual classifiers is 96.11% for ELM type of classifier. There is an improvement of 2.66% when ELM and k-NN classifier fusion takes place with new feature matrix.