Hand gesture recognition using low-budget data glove and cluster-trained probabilistic neural network

Abstract

Purpose – Main purpose is to present methodology which allows efficient hand gesture recognition using low-budget, 5-sensor data glove. To allow widespread use of low-budget data gloves in engineering virtual reality (VR) applications, gesture dictionaries must be enhanced with more ergonomic and symbolically meaningful hand gestures, while providing high gesture recognition rates when used by different seen and unseen users. Design/methodology/approach – The simple boundary-value gesture recognition methodology was replaced by a probabilistic neural network (PNN)-based gesture recognition system able to process simple and complex static gestures. In order to overcome problems inherent to PNN – primarily, slow execution with large training data sets – the proposed gesture recognition system uses clustering ensemble to reduce the training data set without significant deterioration of the quality of training. The reduction of training data set is efficiently performed using three types of clustering algorithms, yielding small number of input vectors that represent the original population very well. Findings – The proposed methodology is capable of providing efficient recognition of simple and complex static gestures and was also successfully tested with gestures of an unseen user, i.e. person who took no part in the training phase. Practical implications – The hand gesture recognition system based on the proposed methodology enables the use of affordable data gloves with a small number of sensors in VR engineering applications which require complex static gestures, including assembly and maintenance simulations. Originality/value – According to literature, there are no similar solutions that allow efficient recognition of simple and complex static hand gestures, based on a 5-sensor data glove.